

small (250x250 max)
medium (500x500 max)
Large
Extra Large
large ( > 500x500)
Full Resolution


ii TAPPLICATION OF DISAGGREGATE DISCRETE CHOICE MODEL FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT By PEERAPOL SITTIVIJAN Master of Engineering Asian Institute of Technology Pathumthani, Thailand 2001 Submitted to the Faculty of the Graduate College of the Oklahoma State University In partial fulfillment of the requirements for the Degree of MASTER OF SCIENCE May, 2009 ii APPLICATION OF DISAGGREGATE DISCRETE CHOICE MODELS FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT Thesis Approved: Dr. Manjunath Kamath___________ Thesis Adviser Dr. Ricki. G. Ingalls______________ Dr. Tieming Liu_________________ Dr. Balabhaskar Balasundaram_____ Dr. A. Gordon Emslie____________ Dean of the Graduate College iii ACKNOWLEDGEMENTS It is difficult to express my gratitude for the amount of help that my advisor, Prof. Dr. Manjunath Kamath, rendered in every respect, for being patient and for giving me an opportunity to carry out this research work as well as giving various valuable suggestions throughout my staying here as a Master's student. I owe a lot to him for providing advice regarding the course selection, moral and financial support and last but not the least for proof reading this thesis. I would also like to thank the committee members, Dr. Ricki Ingalls and Dr. Tieming Liu as well as Dr. Balabhaskar Balasundaram, for providing valuable information and insights into my research. I would also take this opportunity to thank all my friends in Center for Computer Integrated Manufacturing Enterprises (CCiMe) and Center of Engineering Logistics and Distribution (CELDi) for their support at various stages of my thesis. Finally, I would like to dedicate this thesis to my parents, Mr. Vijan Sittivijan, my father, and Mrs. Luksanee Sittivijan, my mother and all other close relatives for their love and support. I owe a lot to them for constantly motivating me with a lot of advice every time I need. iv LIST OF CONTENTS INTRODUCTION………………………………………………………………….……………1 1.1 TAXONOMY OF FREIGHT TRANSPORTATION MODELS………………………..2 1.2 DISAGGREGATE DISCRETE CHOICE MODELS…………………………………...4 1.3 PROBLEMS STATEMENT AND NEED FOR THE RESEARCH……………….…....5 1.4 OUTLINE OF THE THESIS…………………………………………………………….6 LITERATURE REVIEW……………………………………………………………………….8 2.1 REVIEW OF FREIGHT TRANSPORTATION MODELS IN PRACTICE.....………....8 2.2 TRANSPORTATION MODAL SPLIT AND FLOW ASSIGNMENT………………...15 2.3 MULTIMODAL NETWORK MODELS……………………………………………....23 2.4 CLASSIFICATION OF COMMODITIES BASED ON LOGISTICAL REQUIREMENTS……………………………………………………………………....24 2.5 APPLICATIONS OF DISAGGREGATE DISCRETE CHOICE MODELS IN TRANSPORTATION…………………………………………………………………....26 2.6 REVEALED AND STATED PREFERENCE SURVEY FOR PARAMETER .ESTIMATION…………………………………………………………………………..31 STATEMENT OF RESEARCH…..……………………………………………………………35 3.1 RESEARCH OBJECTIVES……..……………………………………………………....35 3.2 RESEARCH SCOPE AND LIMITATION..…………………………………………….36 3.3 RESEARCH CONTRIBUTION……...………………………………...…………..…....37 v RESEARCH APPROACH………..……………………………………………………………38 4.1 EXPLANATION OF THE RESEARCH TASKS ....…….............…………………......39 CRITERIA AND MODELS TO CLASSIFY COMMODITIES INTO LOGISTICAL FAMILIES…………………….………………………………………………..………………..43 5.1 THE CONCEPT OF LOGISTICAL FAMILIES…………………….………………….43 5.2 DISSIMILARITY COEFFICIENTS OF COMMODITIES……………………….….....49 5.3 OPTIMIZATION MODELS FOR CLASSIFICATION………………………………...50 5.4 EXPERIMENTS AND RESULTS OF THE MODELS……………………………...…53 IDENTIFICATION OF EXPLANATORY VARIABLES AND THE STRUCTURE OF THE DISCRETE CHOICE MODEL……………………...…………………………………..58 6.1 VARIABLES BASED ON SUPPLY CHAIN AND LOGISTICS THEORETICAL PERSPECTIVE..........................................................………………..…….……………58 6.2 VARIABLES BASED ON PRACTICAL SHIPPERS AND CARRIERS PERSPECTIVES .........................................................................................….….....…...62 6.3 RELATIONSHIP AMONG ATTRIBUTES...........................................….…..…..…….72 6.4 IDENTIFICATION OF INFLUENTIAL ATTRIBUTES AND MODEL STRUCTURE......75 6.5 MODIFIED UTILITY FUNCTION FOR PATH OVERLAPPING PROBLEM..……...83 ALGORITHM FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT..........................................................................................85 7.1 HEURISTIC ALGORITHM FOR STOCHASTIC USER EQUILIBRIUM FREIGHT NETWORK FLOW ASSIGNMENT................................................................................85 7.2 NUMERICAL EXPERIMENT.........................................................................................90 SUMMARY AND FUTURE RESEARCH....................................…………………………...107 REFERENCES……………………..……………………………………………..…………....109 vi THEORY, DERIVATION AND PROPERTIES OF DISAGGREGATE DISCRETE CHOICE MODEL..………………..…………………………………………………………..122 A1.1 DISCRETE CHOICE AND RANDOM UTILITY THEORY…………………..…...122 A1.2 CHOICE SET DETERMINATION AND FORMATION…………………………...129 A1.3 INDEPENDENCE FROM IRRELEVANT ALTERNATIVES PROPERTY (IIA)…………………………………….....…………………………....130 A1.4 STATE OF THE ART IN DISAGGREGATE DISCRETE CHOICE MODEL RESEARCH ………………………………………….....…………………………...133 FIRST LEVEL (2DIGIT) STCC AND STCG DEFINITION………………………...……136 THE LOGISTICAL CHARACTERISTICS AND DISSIMILARITY COEFFICIENTS...138 A3.1 LOGISTICAL CHARACGTERISTICS OF 8 COMMODITIES.......…………..…...138 A3.2 LOGISTICAL CHARACGTERISTICS OF 49 COMMODITIES....... ………..…...139 CONVENTIONAL MATHEMATICAL PROGRAM FOR STOCHASTIC USER EQUILIBRIUM ASSIGNMENT WITH TRAVEL TIME UTILITY FUNCTION............143 THE DETAILS OF EXPERIMENT........................................................................................154 A5.1 THE DETAILS OF LINK CHARACTERISTICS...........................…………..….....154 A5.2 PATHLINK INCIDENCE MATRIX..............................................…………...…....155 A5.3 TRANSPORTATION COSTS BY MODES.......................................................…....156 A5.4 CALCULATION OF TOTAL UTILITY OF EACH ALTERNATIVE OF MEAT/SEAFOOD...............................................................................................…....157 A5.5 CALCULATION OF TOTAL UTILITY OF EACH ALTERNATIVE OF NATURAL SANDS.............................................................................................…....160 vii LIST OF FIGURES FIGURE 2.1 LINKS FEEDING TRIPS INTO A NODE..........................................................21 FIGURE 2.2 SIMPLE NETWORK OF WATERWAYS..........................................................23 FIGURE 2.3 CORRESPONDING VIRTUAL NETWORK.....................................................23 FIGURE 2.4 DATA ENRISHMENT PARADIGM...................................................................34 FIGURE 4.1 FLOW CHART OF THE THESIS TASKS.........................................................39 FIGURE 5.1 METHODOLOGY FOR COMMODITY CLASSIFICATION IN THE SMILE MODEL......................................................................................45 FIGURE 6.1 RELATIONS AMONG SHIPMENT SIZE, INVENTORY COST AND TRASPORT COSTS..............................................................................................59 FIGURE 6.2 THE INTERACTION AMONG RELATED FACTORS..................................74 FIGURE 6.3 TRANSIT TIME OF A PATH..............................................................................76 FIGURE 6.4 TRANSIT TIME VARIABILITY OF A PATH..................................................76 FIGURE 7.1 INTERMODAL NETWORKS OF HIGHWAY, RAILWAY AND WATERWAY.........................................................................................................92 FIGURE 7.2 SIMPLIFIED INTERMODAL NETWORKS.....................................................92 FIGURE A1.1 AN EXAMPLE OF NESTED LOGIT MODEL FOR RED BUS BLUE BUS SERVICES..................................................................................132 FIGURE A4.1 EXAMPLE NETWORK FOR STOCHASTIC USER EQUILIBRIUM ASSIGNMENT........................................................148 viii LIST OF TABLES TABLE 2.1 SUMMARY OF THE REVIEW OF FREIGHT MODELS IN EUROPE AND THE U.S.............................................................................................................14 TABLE 2.2 SUMMARY OF FREIGHT TRASPORT MODAL SPLIT MODELS...............17 TABLE 2.3 SUMMARY OF ALGORITHMS USED FOR FLOW ASSIGNMENT.............22 TABLE 2.4 LEVELS OF CLASSIFICATION IN STCC.........................................................25 TABLE 2.5 LEVELS OF CLASSIFICATION IN SCTG.........................................................26 TABLE 2.6 CRITERIA FOR PATH FEASIBILITY................................................................29 TABLE 5.1 LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN SMILE MODEL.................................................................................................44 TABLE 5.2 LOGISTICAL FAMILIES IN SAMGODS MODEL...........................................46 TABLE 5.3 LOGISTICAL FAMILIES IN SAMGODS MODEL...........................................46 TABLE 5.4 LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN STEMM...............................................................................................................47 TABLE 5.5 LOGISTICAL FAMILIES IN REDEFINE...........................................................48 TABLE 5.6 MAIN LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN REDEFINE.......................................................................48 TABLE 5.7 LOGISTICAL REQUIREMENTS, VALUES AND WEIGHTS TO CALCULATE DISSIMILARITY COEFFICIENT.............................................50 TABLE 5.8 COMMODITY INDEX AND DESCRIPTION.....................................................55 TABLE 6.1 MAIN SCORE OF FACTORS INFLUENCING TRANSPORT MODE SELECTION OF SHIPPERS AND CARRIERS.................................................65 ix TABLE 6.2 CATEGORY OF ATTRIBUTES AND UNDERLYING TERMS......................68 TABLE 6.3 RELATIVE IMPORTANCE OF TRANSPORT SELECTION CRITERION..69 TABLE 6.4 INFLUENTIAL ATTRIBUTES FOR PORT/FERRY CHOICE........................70 TABLE 6.5 SELECTION CRITERIA FOR TRANSPORTATION MODES........................71 TABLE 6.6 MAIN INFLUENTIAL FACTORS AT MACRO LEVEL..................................72 TABLE 6.7 INFLUENTIAL FACTORS INCLUDED AT THE MACRO LEVEL...............80 TABLE 7.1 DETAILS OF COMMODITIES TRANPORTED AMONG OD PAIRS...........91 TABLE 7.2 DETAILS OF PATH LINKING BETWEEN OD PAIRS....................................94 TABLE 7.3 TRANSPORT COST BY MODE AND SHIPMENT SIZE.................................95 TABLE 7.4 COST OF DAMAGE DURING TRANSIT BY MODE AND COMMODITY...............................................................................................96 TABLE 7.5 COEFFICIENTS OF INDEPENDENT VARIABLES FOR MEAT/SEAFOOD...................................................................................................97 TABLE 7.6 COEFFICIENTS OF INDEPENDENT VARIABLES FOR NATURAL SANDS.................................................................................................97 TABLE 7.7 TOTAL UTILITY OF THE BEST ALTERNATIVES FOR MEAT/SEAFOOD...................................................................................................98 TABLE 7.8 TOTAL UTILITY OF ALTERNATIVES FOR NATURAL SANDS.................99 TABLE 7.9 COMMONALITY FACTORS FOR EACH PATH............................................100 TABLE 7.10 RESULTS OF THE ASSIGNMENT FOR N = 10............................................101 TABLE 7.11 RESULTS OF THE ASSIGNMENT FOR N = 100..........................................102 TABLE 7.12 RESULTS OF THE ASSIGNMENT FOR N = 1000........................................103 TABLE 7.13 ANALYSIS OF RESULTS OF THE ASSIGNMENT FOR N = 10, 100, 1000...................................................................................................106 x TABLE A1.1 DEVELOPMENT OF RANDOM UTILITY THEORY AND ORIGINS OF MNL MODELS................................................................134 TABLE A1.2 APPLICATION OF RANDOM UTILITY THEORY TO TRAVEL DEMAND ANALYSIS.......................................................................................135 TABLE A2.1 FIRST LEVEL (2DIGIT) STCC DEFINITION.............................................136 TABLE A2.2 FIRST LEVEL (2DIGIT) SCTG......................................................................137 1 CHAPTER 1 INTRODUCTION The demand for goods has increased steadily over the past half century and a costeffective freight transportation system has become an integral ingredient of a thriving national economy. Only recently, freight transportation has been systematically analyzed and planned and is comparatively new compared to passenger transportation planning. A critical part of freight transportation analysis and planning is freight transport modeling which is used to forecast and predict behaviors of actors (e.g. freight shippers and carriers) in freight transportation system as well as evaluate related policies and measures. However, the complexity of freight transportation models is far beyond that of passenger transportation. As indicated by Chiang et al. (1980), “In modeling freight transportation systems, models have been developed by researchers from many disciplines using many different approaches in an attempt to solve many different problems. This is just one indication that freight transportation involves complicated decisionmaking processes.” Other factors contributing to the complexity of freight transportation modeling include variables affecting freight movement patterns, for example, locations, range of transported commodities, characteristics and nature of raw materials and end products, manufacturing operations and demand variation and pricing (Ortuzar and Willumsen 1990).  2 In addition, decisionmaking processes in freight transport also relate to many interdependent actors such as suppliers, manufacturers, consumers, shippers (owners of goods who select modes of transport), carriers (owners of transport services who select transport routes) and government as infrastructure providers (Harker 1987 and Tavasszy 1996). 1.1 TAXONOMY OF FREIGHT TRANSPORTATION MODELS Freight transportation models can be generally classified into two main categories: operational models for short to mediumterm decision making and strategic models for longterm decision making. The details and differences between these two types of freight transport models are as follows (Kristiansen and Petersen 2002, Tavasszy et al. 2000). Operational Models – This type of freight transportation models is at the firm level and usually applied for optimization purposes. Compared to strategic models, operational models are closer to actual decision making and more detailed in their description of logistic activities. Besides, a multitude of data is potentially available for modeling at the operational level. Some examples of issues to be analyzed by this type of models are change in cost structure, change in transport market, weight and dimension of vehicle, locations of distribution centers and fleet and crew arrangement. Strategic Models – In contrast to operational models, a strategic model is an aggregation of firmrelated flows. Strategic models are usually descriptive in nature and used to obtain insight into the impact of freight flows on the infrastructure network for longterm planning purposes rather than to optimize decisionmaking processes as required by operational models. Strategic models can be used to analyze, for instance, effects of 3 transport policy measures on longterm patterns and modal distribution, effects of specific transport infrastructure investment projects on traffic pattern/modal distribution and socioeconomic and environmental impacts as well as assessment of transport network development plans. Some examples of policy questions to be analyzed by strategic models are (i) what is the influence of central distribution on transport pattern and mode share, (ii) what is the competition between ports and (iii) what consequences will multimodal transport policy have on utilization of the different transport networks. Most strategic modeling concepts applied in freight transportation have originally been developed based on the conventional fourstep sequential model approach widely applied in passenger transportation. However, the context of fourstep model in freight transportation is quite different from passenger transport as follows; (De Jong et al. 2004) Production and attraction: In this step, the quantities of goods to be produced in various origin zones and the demand for goods that are attracted/consumed in various destination zones are determined (the marginal of origindestination (OD) matrix). The output dimension is tons of goods. In intermediate stages of the production and attraction models, the dimension of freight flows could be converted to monetary units or vehicle units i.e. number of trucks used to transport goods. Distribution: The flows for each commodity type transported between each origin and destination (each cell of the OD matrix) pair are determined. Modal split: The allocation of the commodity flows to modes, for example, highway, rail and waterway. 4 Flow Assignment: Freight flows are assigned onto the network of each mode in this step. Flows could be assigned directly in tons onto the network or converted into vehicleunits before being assigned onto the network. 1.2 DISAGGREGATE DISCRETE CHOICE MODELS Aggregate demand or first generation transport models are either based on observed relations for groups of travelers/actors in the system or on average relations at a zonal level. Disaggregate demand or second generation models are based on observations of individual actors in the system, therefore, enable more realistic models. Spear (1977) summarized some advantages of disaggregate models as follows. Disaggregate models are based on theories of individual behaviors, therefore, an important advantage of the disaggregate models over conventional aggregate ones is that it is more likely that disaggregate models are stable and transferable in time and space. Since disaggregate models are based on individual data, they require fewer data points and are less likely to suffer from biases due to correlation between aggregate units. Disaggregate models are stochastic and they yield the probability of choosing each alternative rather than indicate which one is selected. Disaggregate models can have explicitly estimated coefficients and allow any numbers and specification of the explanatory variables while generalized cost function in aggregate models is generally limited to only costrelated variables (e.g. travel cost and travel time) and fixed parameters. This implies that disaggregate 5 models are a more flexible representation of policy variables. Besides, coefficients of explanatory variables directly reflect the relative importance of each attribute. The details of disaggregate discrete choice theory, derivation of models and some important properties are described in Appendix A1. 1.3 PROBLEM STATEMENT AND NEED FOR THE RESEARCH Freight transportation activities in the present circumstances are closely related to a larger context of logistics decisionmaking (e.g. inventory policy, warehouse and distribution center locations). Nowadays, commercial companies do not only consider transportation as the immediate cost of moving goods from one place to another, but they view transportation process as a part of the whole logistic concept. According to this concept, capital requirements related to easy and fast market access might be more important than the direct transport costs. Firms are also aware of the importance of response times and reliable delivery. Nevertheless, this logistic requirement is not relevant to all types of goods. For example, bulk products like crude oil and coal are commodities that can be kept in storage for relatively long periods. In this case, the direct transport costs are still of major importance for the choice of mode and transport route (Kristiansen and Petersen 2002). Most existing freight transport demand models used for mode and route selection still focus only on the direct factors e.g. transport cost and transit time. However, in practice, as described previously, other important logistics factors could also affect the transportation decisions of shippers and carriers. Therefore, this thesis effort identifies these factors and studies how they can be incorporated into freight transport demand models in order to improve the forecasting capability of the model. Conceptually, one 6 possible way to do so is by the application of discrete choice approach to mode and route selection models. Although, it is possible in principle, based on the review of freight transportation models both in the U.S. and in Europe and a study by De Jong et al. (2004), disaggregate behavioral models are not common in freight transportation models. Commodity characteristics are a key factor in decision making relating to transport mode and route selection. Further, with the application of the discrete choice model, commodities in the same group are assumed to have similar transportation behaviors. Therefore, a part of this research is also devoted to the classification of commodity groups. 1.4 OUTLINE OF THE THESIS The rest of this thesis document is organized as follows. Chapter 2 includes a review of freight transportation models in practice, modal split and flow assignment models, multimodal network models, classification of commodity groups based on logistical requirements, application of discrete choice models in transport mode and route selection, and revealed and stated preference survey for parameter estimation. This is followed by Chapter 3, the research statement, which explains the research goals, objectives, scope and limitations and contributions to the field of freight transportation demand modeling. Chapter 4 presents the research approach and enumerates the steps in order to accomplish the objectives. Chapter 5 is devoted to the review and reclassification of STCC and STCG commodity groups based on commodity characteristics and logistical requirements. The mathematical models based on the network clustering concept are used to classify commodities into logistical families. Chapter 6 presents potential factors, besides direct transport cost and transit time, for freight transport mode 7 and route selection model. The structure of discrete choice model based on the identified potential factors for freight transport mode and route selection is also proposed in the chapter. Finally, in Chapter 7, the deployment of the proposed discrete choice model for multicommodities multimodal stochastic network flow assignment is explored. First, the conventional stochastic user equilibrium assignment model is studied. Then, the application of the proposed discrete choice model to the conventional assignment model is presented. A heuristic algorithm to assign freight flows onto multimodal stochastic and congested network based on the proposed discrete choice model is developed. A proofof concept implementation of the heuristic algorithm for a simplified network with multiple origindestination pairs and commodities is also presented. 8 CHAPTER 2 LITERATURE REVIEW This chapter begins with the review of freight transportation models developed both in the U.S. and Europe. Then, algorithms used for transportation modal split and flow assignment are reviewed. In section 2.3, multimodal network models and the methods to develop such models are described. The concept of classifying commodities according to their logistical requirements is presented in section 2.4. The application of disaggregate discrete choice models in the transportation context, especially in mode and route selection is summarized in section 2.5. The chapter ends with a review of revealed and stated preference survey for parameter estimation in discrete choice models. 2.1 REVIEW OF FREIGHT TRANSPORTATION MODELS IN PRACTICE In this section, freight transportation models developed both in the U.S. and Europe are reviewed. 2.1.1 Freight Transport Models in Europe 2.1.1.1 Strategic Model for Integrated Logistic Evaluation (SMILE) – SMILE is a strategic model for freight transport and logistics used in Netherlands. 9 Although it is a strategic model describing aggregate level of national freight transport, it has the ability to incorporate the relationship among production, inventory and transportation at disaggregate level. This relationship can be represented and analyzed in the model by a threelevel modeling approach as follows; (Tavasszy et al. 1998) Module of production, sales and sourcing – the location pattern of both production and consumption are established by using Make/Use tables. After the volume and nature of production and consumption at different locations are determined, the spatial distribution from production sites to consumption sites resulting from comparative price differences and the resistance of geographical, organizational and institutional differences are determined. Inventory Module – The main function of this secondlevel model is to link trade relations to transport relations by considering warehousing services. New groups of commodities are formed based on the product and market characteristics such as the value of products per cubic meter, packing density and perishability and so forth. Optimal locations of distribution centers are determined by using criteria such as lead time, closeness to activity centers and available modes of transport. Then conditional on the locations of distribution centers, multinomial logit models are used to assign flows to alternative channel types. Transportation Module – Freight flows in each distribution channel resulting from the Inventory Module are assigned onto the multimodal network of six modes using aggregate approach to find the optimal route (AllorNothing) with cost minimization. 2.1.1.2 Network Model for Norwegian Freight Transportation (NEMO) – The conventional fourstep model including generation, distribution, modal split and 10 assignment is applied directly in NEMO. Only eleven groups of products are included in the model. The network in the model is a multimodal network, therefore, allowing for simultaneous mode and route analysis. The transport modes included in the network are road, rail, seaborne transport, air freight and pipeline. The assignment algorithm in the model is aggregate allornothing based on cost minimization. The cost structure is comprised of reload costs when mode transfer occurs, qualitative costs (e.g. risk of delay and travel time) and truck, ship and rail operation costs (Hovi and Vold 2003). 2.1.1.3 The Walloon Freight Transport Model for Belgium (WFTM) – WFTM also applies the conventional fourstep model. The base year OD matrices in 1995 of ten commodity classes were constructed as the part of freight flow generation and distribution steps. In this model, modal split and flow assignment are also simultaneously analyzed on its multimodal network using the system optimal algorithm (e.g. an assignment algorithm aiming to minimize the total generalized cost in the network). An interesting concept of WFTM is the utilization of a virtual network to represent multimodal network in physical configurations and operational characteristics. The operational characteristics modeled in WFTM are loading, transshipment, waiting and handling (Jourquin and Beuthe 2000 and Jourquin and Limbourg 2006). The details of the virtual network will be described in more detail again in section 2.3 of the literature review. 2.1.1.4 National Freight Model System for Sweden (SAMGODS)  Base year OD matrices in the year 1997 were constructed for six main groups, split into bulk/general, cargo, high/low density and high/low value. The basis for the model is a set of inputoutput tables. The model calculates forecasts of production, import, export, inputs to production and consumption in monetary units for 31 different sectors. For the network 11 model in SAMGODS, a multimodal network of road and rail modes is used and the flow assignment is analyzed by the system optimal approach which minimizes the generalized cost of the whole network (MEP and SWP 2002 and De Jong et al. 2002). 2.1.1.5 Decision Support System for Transport Policies of Italy – Here also the concept of the fourstep modeling approach is used. Generation and distribution models are represented by multiregional input/output models. In this Italian freight transport model, modal split is separated from the assignment step. Disaggregate discrete choice model is used to forecast mode share among seven modes: road by own truck, truck of carrier for a single shipment or truck of carrier under contract, traditional rail, combined rail (contained, swap bodies or semitrailers), shipper by road, shipper by rail, shipper with mode chosen by himself. The segmentation of commodities believed to influence mode choice such as perishable, consumer and capital goods are classified. The assignment stage takes the multimodal OD matrices and assigns them to networks. For road, random utility path choice models are used (MEP and SWP 2002). 2.1.2 Freight Transport Models in the U.S. For the statewide freight transportation models in the U.S., the structures of the models are quite similar. The unit of flow is typically number of trucks for vehicular flows and tons for commodity flows. The main sources of OD data of freight are Commodity Flow Survey (CFS) (Bureau of the Census and Bureau of Transportation Statistics, 1997) and TRANSEARCH database (http://www.globalinsight.com/Transearch). Most of the models use the fourstep sequential modelling approach. Details of the models for some states are briefly described below; 12 2.1.2.1 Statewide Freight Transport Model for Indiana – There are 145 Traffic Analysis Zones (TAZs) in the model representing 92 counties in Indiana and 53 more for other 47 states (not including Hawaii and Alaska) in the U.S. CFS (Bureau of the Census and Bureau of Transportation Statistics, 1997) is used as the main source for the OD flow matrix. Two modes of transport, namely, truck and rail, are considered in the model. The modal split model uses the CFS modesplit ratio in 1993 for the future years. The commodity flows in tons for both rail and truck are converted to vehicle trips and assigned onto their own network by AllorNothing approach (Horowitz and Farmer 1999). 2.1.2.2 Statewide Freight Transport Model for Wisconsin – There are 132 Traffic Analysis Zones (TAZs) in the model representing 72 counties in Wisconsin and 60 more for other 47 states in the TRANSEARCH database is used as the main source for OD flow matrix. Highway, rail, waterway and air transport modes are considered in this model. Aggregate logit model for all commodities is used to split mode share. Based on the information from vehicle inventory and usage survey, payload factors are used to convert commodity tons to truck trips and conversion factors are used to convert annual truck traffic to daily volumes. Daily truck volumes are assigned onto the network by using multiclass User Equilibrium assignment with preloaded passenger volumes. However, commodity flows in tons are assigned onto the rail network by using Allor Nothing algorithm (Center for Urban Transportation Studies, 1999 and Transportation Systems Design Graduate Students, 2006). 2.1.2.3 Statewide Freight Transport Model for Iowa – There are 145 TAZs in the Iowa statewide freight model representing 99 counties in Iowa and Business Economic Areas 13 (BEAs) in other states. Only three main commodity groups (e.g. grain, meat products and machinery) are considered important and included in the model. The main source of OD freight flows is TRANSEARCH. Highway transport by trucks is only the transport mode considered in the model. Commodity flows in tons are assigned onto the network using AllorNothing algorithm (Souleyrette et al. 1996). 2.1.2.4 Statewide Freight Transport Model for Oklahoma – There are 204 TAZs in the model representing 77 counties in Oklahoma and 127 business areas in other states. The main source of OD freight flows is the Freight Analysis Framework (FAF2.2) developed by Federal Highway Administration, US Department of Transport (Federal Highway Administration, 2006). The total commodities are classified into 43 groups based on Standard Classification of Transported Goods (SCTG). The ratios among transportation modes; highway, railway and waterway in 2002 are used for mode choice in the future years. The commodity flows transported by railway and waterway are assigned on their own network using shortest travel distance algorithm. For highway mode, the flows are assigned onto the network by using shortest travel distance, shortest travel time, user equilibrium and system optimal algorithms. Furthermore, a mathematical programming model is underdevelopment in order to assign highway freight flows to minimize the total system travel time by simplifying the travel timeflow relation to be piecewise linear. All capacitated assignments are analyzed after preloading passenger flows onto the network (Ingalls et al. 2007 and Ingalls et al. 2009) A summary of the reviews of the above freight models, especially with respect to the aspects of modal split, flow assignment, multimodal consideration and inclusion of logistical requirements is shown in Table 2.1. 14 Table 2.1: Summary of the Review of Freight Models in Europe and the U.S. Model Name Modal Split Flow Assignment Multimodal Consideration Logistical Requirements Consideration Strategic Model for Integrated Logistics Evaluation (SMILE) Not included Not included Multimodal network assignment by optimal route (AllorNothing) algorithm based on cost minimization Included in the model based on commodity classication and inventory module considering optimal distribution channel Norwegian National Freight Model System (Nemo) Not included Not included Multimodal network assignment by optimal route (AllorNothing) algorithm based on cost minimization Not included The Walloon Freight Transport Model for Belgium (WFTM) Not included Not included Multimodal network assignment by System Optimal to minimize costs of the total network Included in the model based on virtual network considering logistical activities at transfer/transshipment facilities Swedish National Freigt Models System (SAMGODS) Not included Not included Multimodal network assignment by System Optimal to minimize costs of the total network Included in the model based on commodity classication as bulk/general, cargo, high/low density and high/low value Decision Support System for Transport Policies of Italy Disaggregate discrete choice model Random utility path choice model for road transport Not included Included by segmentation of commodities believed to influence mode choice: perishable, consumer and capital goods Statewide Freight Tranport Model for Indiana Ratio of mode share from CFS database Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Tranport Model for Wisconsin Aggregate logit model Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Tranport Model for Iowa Only truck mode is considered Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Transport Model for Oklahoma Ratio of mode share from FAF database for highway, railway and waterway modes Railway and waterway flows by shortest distance and capacitated assignment for highway Not included Not included 1. Freight Transport Models in Europe 2. Statewide Freight Transport Models in U.S. 15 2.2 TRANSPORTATION MODAL SPLIT AND FLOW ASSIGNMENT As described in Chapter 1, modal split and flow assignment models are the last two steps within the conventional fourstep model. The main functions of modal split and flow assignment models, in the context of freight transportation, are to allocate commodity flows to different modes and assign flow in each mode onto the corresponding transportation network. 2.2.1 Modal Split models – Many algorithms have been used to represent decisionmaking for mode selection ranging from an easy and coarse method to a more complicated and detailed one. The easiest method is to use empirical data to develop the ratios of sharing among modes. Another possible way is to construct a diversion curve describing the relationship between proportions of flows by mode against the difference of some factors (e.g. cost and time difference) between any two modes (Ortuzar and Willumsen 1990). De Jong et al. (2004) also provided a taxonomy of modal split models as follows. Elasticitybased models reflect the effect of changing a single variable (e.g. cost of some modes) to the proportion of mode selection. This type of model is usually used for strategic evaluation or for a quick first approximation. Elasticities can be derived from other models such as aggregate mode choice model or from expert knowledge. Aggregate modal choice model – The models are generated from the concept of entropymaximization. In entropymaximization, a system is made up of a large number of distinct elements and can be classified into three main levels: micro, meso and macro. The concept postulates that there are numerous and different states in micro and meso levels which produce the same state in macro level and assumes that 16 all micro states consistent with our information about macro states are equally likely to occur. The derived aggregate models are similar to disaggregate discrete choice models in logit family. However, explanatory variables are at the aggregate level and parameters are estimated based on zonal/interzonal information. The main disadvantage of aggregate modal choice model is the lack of insight into causality of mode choice from the individual perspectives. Neoclassical models – the models are derived from the economic theory of the firm. For a cost function, transport services are considered as one of the inputs. The explanatory variable in the models is the budget share of some modes in the total cost. This type of model is difficult to combine with the conventional fourstep model because the share in the transport volume is the relevant variable. Direct demand models – These models can generate the number of trips by modes directly (unlike generated in market share forms by other types of models). Because the models directly generate number of trips by each mode, it is quite separate from the product/attraction and distribution steps in the fourstep model and make it difficult to incorporate it into the fourstep framework. 17 Disaggregate modal choice models – This type of models is derived from the concept of random utility theory. The model structures of logit family are similar to those of aggregate modal choice models. To estimate models’ parameters, behavioral data from individual units, e.g. shipper and carrier in the freight transport context, are required. More details of disaggregate modal choice models are presented in Appendix A1 and section 2.5 in this chapter. A summary of freight transport modal split models is presented in Table 2.2 (De Jong et al. 2004). 2.2.2 Flow Assignment Models – Flow assignment models can be classified with respect to supply characteristics and demand assumptions underlying the models. For supply characteristics, a network can be considered congested when link performances (e.g. travel time and cost) depend on link flows or uncongested, if link performances are Type of Model Advantages Disadvantages Modal share ratios/Diversion curve Quickest and easiest to develop based on historical data Lack of supporting theory and concerns about predicting capability Elasticitybased models Very limited data requirements and quick to develop Elasticities may not be transferable; measure only impact of single variable; no synergies Aggregate modal split models Limited data requirements Little insight into causality based on real behaviors of individuals and limited scope for policy effects Neoclassical models Limited data requirements and supported by theoretical basis Hard to integrate into fourstep models Disaggregate modal split models Supported by theoretical basis, able to include casual variables and policy sensitive measures based on real behaviors of individuals Need disaggregate data from individual (shippers/carriers) for parameter estimation Table 2.2: Summary of Freight Transport Modal Split Models Source: National and International Freight Transport Models: An Overview and Ideas for Future Development by De Jong et al. (2004) 18 fixed and independent of link flows. For a congested network, costflow or travel time flow curves are used to represent the relationship between link flows and link performance. Costflow relationships proposed by researchers are as follows; Smock (1962) for Detroit Study; exp (V/C) 0 t = t Overgaard (1967) (V/C) 0 t = t Bureau of Public Roads (1964) in the U.S. ] [1 (V/C) 0 t = t + where V is link volume in passenger car equivalent unit/hour. C is link capacity in passenger car equivalent unit/hour t is congested travel time of the link 0 t is freeflow travel time of the link and b are parameters On the demand side, if it is assumed in the model that all users perceive network performances in the same way as unique values and have no personal preference, the model can be considered deterministic. If the model allows the possibility of perceiving network performances differently and users can have personal preference, the model is stochastic. In the case of deterministic assignment, all users of the link are assumed to perceive the link cost in same way as the mean link cost. However, in the case of stochastic assignment, users of a link are allowed to differently perceive the link cost based on the assumption of link cost distribution. 19 In addition, flow assignment models can be aggregate if they are derived from zonal/interzonal data or disaggregate if they are derived based on individual data. Some algorithms developed for flow assignment models are as follows (Ortuzar and Willumsen 1990); Allor–Nothing  This is the simplest route choice and assignment method. The algorithm is for deterministic uncongested network flow assignment. The algorithm is probably reasonable for sparse and uncongested networks e.g. intercity/interstate networks. However, for networks in urban areas where there are high congestion effects, this model may not be realistic. User Equilibrium – This model can be considered deterministic congested network assignment. The model is based on the Wardrop’s first principle (1952) postulated that “Under equilibrium conditions traffic arranges itself in congested networks in such a way that no individual trip maker can reduce his/her path costs by switching routes and all used routes have equal and minimum costs while all unused routes have greater or equal costs”. It means that this algorithm minimizes generalized costs of each individual user in the system. System Optimal – The model is based on the Wardrop’s second principle (1952) postulated that “Under social equilibrium conditions, traffic should be arranged in congested network in such a way that the generalized costs of the whole system are minimized”. In contrast to the user equilibrium reflecting behaviors of individuals trying to minimize their own costs, this algorithm is oriented towards transport planners who try to minimize total system costs. 20 Simulationbased Methods – This approach was proposed by Burrell (1968). It can be considered as an approach for stochastic uncongested network assignment. In this approach, for each link in a network, link cost is separated into objective or engineering costs as measured by a modeler or the observer of the system and subjective cost as perceived by real users in the system. Therefore, it can be assumed that users perceive link cost as random variable with the engineering cost as the mean and distributed according to some known distributional form e.g. Uniform or Normal. To assign flows onto the network, total flows are divided into N segments, Monte Carlo simulation used to generate random generalized cost for each segment, then, flows in each segment are assigned onto the network based on AllorNothing approach. Singlepath Proportional Stochastic Methods – This is another stochastic uncongested network assignment algorithm proposed by Dial (1971). The method is based on a loading algorithm which splits trips arriving at a node between all possible exit nodes. For example, in Figure 2.1, flows from origin I to node B will be split into the flows from Ai to B by splitting factors which can be derived from the extra cost incurred in traveling from the origin to node B via node Ai rather than only via the minimum cost route. 21 Discrete Path Choice Models – There are some applications of discrete choice models in the context of transportation route choice analysis. The applications usually are for stochastic uncongested network assignment. This type of models is discussed in more detail in section 2.5. Stochastic User Equilibrium – In this case, the transportation network is considered for both congested and stochastic aspects. Performances of a link depend on link’s flows and can be perceived differently by each user on the links. The assignment problem can be formulated as a mathematical program and solved to optimality using an algorithm called the Method of Successive Averages (MSA) (Sheffi, 1985). The details and algorithm for stochastic user equilibrium are discussed in Chapter 7. A summary of the algorithms used in flow assignment according to supply and demand characteristics is shown in Table 2.3. I J A4 A2 A1 A3 A5 B Figure 2.1 Links Feeding Trips into a Node Source: Modelling Transport by Ortuzar and Willumsen (1990) 22 Table 2.3: Summary of Algorithms used for Flow Assignment Uncongested Network Congested Network Deterministic AllorNothing User Equilibrium, Systetm Optimal Stochastic Simulationbased, Singlepath proportional, Stochasitic Methods and Discrete Choice Model Sotchastic User Equilibrium 2.3 MULTIMODAL NETWORK MODELS Recently, multimodal transportation has become increasingly important since it can provide greater efficiency and cost savings than monomodal transport. According to the 1997 Commodity Flow Survey (CFS) (Bureau of the Census and Bureau of Transportation Statistics, 1997), tens of thousands of intermodal shipments were operated out of the total shipments of around five million tons (Bureau of the Census and Bureau of Transportation Statistics 1997). Southworth and Peterson (2000) proposed a method to construct a multimodal network from monomodal networks. In their work, three main transportation networks; truck, rail and waterways, are combined together into an intermodal network. Transfer terminals among modes (e.g. truckrail, truckwaterways and waterwaysrail) are used as the connecting nodes in the intermodal network. The detailed National Intermodal Terminals Database (Middendorf 1998) is used for this purpose. Harker (1987), Crainic et al. (1990), Jourquin and Beuthe (2000) and Jourquin and Limbourg (2006) proposed the concept of virtual links to represent specific costs for particular uses of transportation infrastructure because a simple network does not provide an adequate basis for detailed analyses of transport and logistics operations. For example, trucks of different sizes and operating costs can use the same highway or at a terminal, a truck’s load can be transshipped on a train, bundled with some 23 others on a boat or simply unloaded as it reaches its station. Although these operations in the example use the same infrastructure, but the costs of each operation are different. Therefore, virtual links are needed for representing different operations/costs on the same infrastructure. An example of virtual network construction can be illustrated by the following example. Figure 2.2 shows a simple network of waterways consisting of 4 nodes and 3 links. Each link represents waterways that can support ships of 300, 1,350 and 600 tons respectively. Figure 2.3 shows the corresponding virtual network in which alternative operations are enumerated, for example, the shipment can be directly transported by 300ton ships from node a to node d or be transshipped at node b and c then transported to node d. This concept can be very useful for reflecting and integrating detailed logistical operations in freight transportation models. a1 b4 b3 c3 b2 c2 d2 c1 c4 c6 300t b1 b5 c5 d1 1350t 1000t 600t 600t 300t 300t Figure 2.3 Corresponding Virtual Network Source: Multimodal Freight Networks Analysis with NODUS: A Survey of Several Applications by Jourquin and Beuthe (2000). Figure 2.2 Simple Network of Waterways Source: Multimodal Freight Networks Analysis with NODUS: A Survey of Several Applications by Jourquin and Beuthe (2000). a b d 300t 1350t 600t c 24 2.4 CLASSIFICATION OF COMMODITIES BASED ON LOGISTICAL REQUIREMENTS To integrate logistical processes into freight transportation modeling, commodity classification is considered an important issue (REDEFINE 1999 and Tavasszy et al. 1998). This is especially true for mode and route selection steps. Basically, the classification should be based on handling characteristics and the fact that different commodity groups have different values of time. In project REDEFINE and SMILE, commodity groups are classified based on product characteristics and logistical requirements as follows; Bulk or general cargo Density (Kg/m3) Packaging density (packs/unit/ m3) Value (Euro/kg.) Use of distribution centers (Yes/No) Consignment size (small/medium/large) Value of time (low, medium, high) Demand frequency In the U.S., there are two main systems for classification of transported commodities: Standard Transportation Commodity Code (STCC) and Standard Classification of Transported Goods (SCTG). The details for each system are as follows (Federal Highway Administration 2006); Standard Transportation Commodity Code (STCC) – STCC system was developed in 1960s by a special committee of the Association of American Railroads (AAR). The 25 main purpose of the development was to serve users of AAR, particularly, North American Freight Railroads. The annual Railroad Waybill data, 1993 Commodity Flow Survey (CFS), and the first generation of the FAF all used the STCC coding system. There are 4 levels in the hierarchical system from 2digit to 5digit codes. Generally, the first four digits of the STCC represent the industry that produces the commodity, based on the Standard Industrial Classification (SIC) system. The fifth digit of STCC provides product classes within the producing industries. The last two digits of the STCC add commodity detail of particular interest to the railroads. A summary of the various 5digit STCC levels is presented in Table 2.4. The top level (2digit) STCC codes are listed in Appendix A2. Table 2.4: Levels of Classification in STCC Source: Report 4 (R4): FAF Commodity Classification by Federal Highway Administration (2006) Standard Classification of Transported Goods (SCTG)  The U.S. Department of Transportation, U. S. Bureau of the Census, Statistics Canada, and Transport Canada developed the SCTG to replace the STCC for the 1997 and subsequent CFS. The structure of SCTG is similar to that of STCC. There are also 5 levels of commodity codes in SCTG. At the most aggregated level (i.e., 2digit), the SCTG was designed to provide analytically useful commodity groupings for users that are interested in an overview of transported goods. With a small number of exceptions, categories in the 3digit level were designed to include goods for which significant product movements are expected to be recorded in both the United States and Canada. The 4digit SCTG Level Number of Categories Grouping (example) 2digit 37 Major industry classes (01Farm products) 3digit 182 Minor industry classes (012Fresh fruits or tree nuts) 4digit 444 Specific industries (0121Citrus fruits 5digit 1,202 Product classes (01214Oranges) 26 Level of Hierarchy Number of Categories Grouping (example) 2digit 42 Analytical overview 3digit 133 U.S.Canadian product groups 4digit 283 Transportation characteristics 5digit 504 CFS 2002 collection level categories were created to reflect industry patterns and transportation characteristics. The most detailed SCTG category, which is at 5digit level, is the collection level for the CFS. At this level, each category was designed to capture significant details that reflect industry patterns and transportation characteristics. Because most 4 and 5 digit SCTG categories primarily contain the products of only one industry, they can be associated with the SIC, as well as with the NAICS. This feature allows comparisons to be conducted with industry data, as well as other SICbased classifications such as the STCC system. The number of categories in each level of SCTG, as used in the 2002 CFS, is summarized in Table 2.5. The first level categories of the SCTG are listed in Appendix A2. 2.5 APPLICATIONS OF DISAGGREGATE DISCRETE CHOICE MODELS IN TRANSPORTATION The early applications of disaggregate discrete choice models in the area of transportation were made mainly for the binary choice (e.g. only two alternatives in a choice set) of travel mode. Some of these studies focused on the tradeoff between travel time and travel cost implied by travel demand models. Thereafter, further development of disaggregate discrete choice models in transportation was directed toward the choice of transport modes with more than two alternatives (multinomial discrete choice models) Source: Report 4 (R4): FAF Commodity Classification by Federal Highway Administration (2006) Table 2.5: Levels of Classifications in SCTG 27 and other transportrelated choice situations such as trip destination, trip frequency, car ownership, residential location and housing and routes selection (CRA 1972, BenAkiva 1973 and 1974, Brand and Manheim 1973, Richards and BenAkiva 1975 and Lerman and BenAkiva 1975.) For mode selection, Cascetta (2001) defined disaggregate mode choice models as the models used to simulate the probabilities of a user to select transport modes from an origin zone to a destination zone. Identification of modal alternatives in choice set depends on each transportation system under study. By using different heuristic approaches, trivially nonavailable alternatives can be eliminated such as choice of walking or bicycling can be excluded from the choice set of interurban transport system. For conditionally nonavailable alternatives, probability of including such alternatives in the choice set can be calculated, then probability of selecting alternatives in choice set can be identified as the conditional probability. Explanatory variables included in utility functions of mode choice models usually are levelofservice/performance attributes of each alternative and individual’s socioeconomic attributes. Examples of performance attributes in mode selection models could be travel time, travel cost, regularity of services and number of transfers. Socioeconomic attributes are characteristics of the decisionmaker in the system. This type of attributes is generic and not dependent on alternatives. Examples of socioeconomic attributes for mode choice models could be gender, age, family income and car ownership. Disaggregate mode choice models are quite common in transportation planning. Examples of passenger transport mode choice models are Warner 1962, Lisco 1967, Quarmby 1967, Lave 1969, Watson 1974, Rassam et al. 1971, 28 Ben Akiva 1973 and McFadden 1974. For freight transport mode choice model, some studies related disaggregate mode choice model are as follows (De Jong et al. 2004); Winston (1981): probit model for the choice between road and rail transport by commodity group in the U.S. Jiang et al. (1999): nested logit model on the French 1988 shippers survey. Nuzzolo and Russo (1995): mode choice model for the Italian national model. Fosgerau (1996): mode choice model on revealed and stated preference data. Reynaud and Jiang (2000): European freight model focusing on operating systems for rail developed with mode choice model on revealed and stated preference. Chiang et al. (1981): disaggregate model for selection of supplier, shipment size and transport mode. McFadden et al. (1985): disaggregate model for shipment and mode selection. Jovicic (1996): SP and RP interviews of Danish freight shippers and carriers to examine the importance of a number of different parameters describing their transport decision making. For path selection, disaggregate path choice models provide the probability of path selection of a user from an origin zone to a destination zone for a particular transport mode (Cascetta 2001). Disaggregate path choice models can be typified into two main categories: pretrip choice where the whole traveling path is chosen before starting the trip and pretrip/en route mixed choice where the route is chosen both before the starting and during the trip, for example, the selection of route for urban transit system with high frequency and low regularity. As mentioned in Appendix A1, the crucial parts of 29 disaggregate path choice models are identification and inclusion of path alternatives into the choice set. There are two ways to handle this issue as follows (Cascetta 2001): Exhaustive approach – The approach includes all elementary paths in the network into the choice set. It may generate a very large number of routes sharing many links. Selective approach – In contrast to the exhaustive approach, this approach identifies only some elementary paths in the choice set on the basis of heuristic rules. The heuristic rules used in path choice models depend on the application context. Some examples of heuristic rules for path choice alternatives are shown in Table 2.6. Another important issue for disaggregate path choice model development is the correlation among alternatives. It can be seen that alternatives in route choice models could be highly correlated because there are many alternative routes sharing links. In the case that the random part of the utility function is assumed to be Gumbel distribution and the discrete choice model is called "logit choice model", this correlation issue could introduce complexity (the derivation of logit choice model and its property are presented Table 2.6: Criteria for Path Feasibility Selection Criteria Specification Topological A path is feasible (Dial efficient) if each link "goes away" from the origin and/or "move towards" the destination. Comparison of costs Paths with a generalized cost not exceeding by more than X times of minimum cost. Progressive The first n minimum generalized cost paths. Multiattribute Minimum paths with respect to various attributes (usually the relevant perfomance variables such as travel time, monetary cost, motorway distance and so forth). Behavioral Paths excluding behaviorally unrealistic link sequences (e.g. repeated entrances and exits for the same motorway). Distinctive Paths overlapping for no more than a given percentage of their length. Source: Transportation Systems Engineering: Theory and Methods by Cascetta (2001) 30 in Appendix A1). The reason is that there is a key assumption to derive the logit choice models that the disturbance term in the utility function is identically and independently distributed meaning that all alternatives must be independent. The explanatory variables generally applied in path choice models are path performance or levelofservice such as travel time, travel cost, number of transfers, service frequency, level of congestion and so forth. Disaggregate path choice models are not widely applied in passenger transport and, according to the literature review, they are not common in freight transport either. Limited application of disaggregate path choice model may stem from their complexity and the difficulty in data collection compared to the aggregate model. Examples of path choice models in passenger transport are as follows; BenAkiva et al. (1984): Discrete path choice model for passenger transport by defining choice sets of labeled paths to transform a large number of physical routes into a smaller number of labeled routes. Dial (1971): A Probabilistic Multipath Traffic Assignment Model approach to include reasonable paths into the choice set. Paths included in the choice set should be composed of links that would not move the traveler farther away from his/her destination. Burrell (1968): Stochastic models seeking to account for variations in drivers’ perceptions of travel times or costs by means of a probability distribution for perceived link performances. Hidano (1983): Proposing individuals’ route plan in a hierarchical fashion starting at the lowest level in the road network, proceeding up a hierarchy and then going down the hierarchy in the vicinity of their destination. 31 2.6 REVEALED AND STATED PREFERENCE SURVEY FOR PARAMETER ESTIMATION Disaggregate discrete choice models are data dependent meaning that the quality and accuracy of the models largely depend on the data used for the models’ parameters estimation. Before the mid of 1980s, parameter estimation of discrete choice models was dominated by revealed preference data. Revealed preference (RP) data is based on real behavior observed in an actual system. However, revealed preference data has some limitations as follows (Ortuzar and Willumsen 1990); Observations of actual choices may not provide sufficient variability for constructing good models for evaluation and forecasting. For example, travel time and travel cost may be highly correlated and may be very difficult to separate their effects in model estimation. The observed behavior may be dominated by a few factors making it very difficult to detect the relative importance of other variables, especially for qualitative variables. For the alternatives which are entirely new such as completely new mode of transport, it is very difficult to collect data by using revealed preference technique. According to these reasons, a new technique was required to fill the gaps. By the end of 1970s, stated preference (SP) technique that originated in the field of market research was proposed to overcome the limitations of revealed preference data. The SP technique offers a way of experimenting with choices directly, thus solving some of the above limitations. SP technique estimates the parameters of discrete choice model based on an analysis of the response to hypothetical choices which can cover a wider range of attributes and conditions than the real behavior as applied in RP technique. By using SP 32 survey, individuals are asked about which choice they would make in one or more hypothetical situations. Cascetta (2001) summarized the advantages of SP over RP technique as follows; SP technique allows the introduction of choice alternatives not available in the present. They can control the variation of relevant attributes outside the present range to obtain better estimates of the relative coefficient. They can introduce new attributes and their coefficients, especially, qualitative ones such as convenience and safety. However, these advantages come with the tradeoff of introducing some distortion in the results because individuals respond based only on the hypothetical and not the real situation. Besides, it is possible that the stated choices are presented in unrealistic ways for example, some attributes presented to individuals might be missing or there may be fatigue and justification bias effects and these situations can further increase the distortion in the models. Hensher (1994) summarized the steps in implementing SP survey as follows; Task 1: Identification of the set of attributes – There may be a large number of attributes to be included in an alternative, therefore, it is required to decide early on which attributes should be included. Task 2: Selection of the measurement unit for each attribute – In most cases the units of attributes are unambiguous, for example, hours for travel time and dollars for travel cost. However, for some qualitative attributes such as convenience, reliability and safety, the units of attributes should be clearly identified to avoid possible confusion of respondents. 33 Task 3: Specification of the number and magnitudes of attributes – The levels of attributes are specified in this step to measure variation of each attribute in SP survey. For example, the levels of travel cost can be divided into three levels: high, medium and low cost. The value of attributes in each level can be identified based on the existing value in an actual system. Task 4: Statistical Design– In this step, attributes and their levels are combined to construct hypothetical situations. The design of these hypothetical situations can be based upon the experimental design principle widely used in statistics. Full factorial or fractional factorial design can be applied. Task 5: Translation of statistical design into a set of questions for data collection– The experimental design in task 4 is translated into a set of question to ask respondents. One should be careful about setting up the questions used to asked respondents. McFadden and Leonard (1992) conducted the tests of SP methods and compared results from alternative SP experiments different in response format, question phrasing and information provided to the respondent. They found great sensitivity of the results according to these factors. As mentioned earlier, both RP and SP techniques have their own advantages and disadvantages. RP data provides the real behaviors of individuals in an actual situation while SP data allows the introduction of nonexisting alternatives and control of attributes’ variation. Therefore, to improve the models’ estimation, one may be required to combine both RP and SP data together. This process was originally proposed by Morikawa (1989). The data enrichment paradigm (Louviere et al. 2000) according to Morikawa’s concept can be depicted in Figure 2.4. For this paradigm, RP data are viewed 34 as the standard of comparison among alternatives and SP data are seen as useful only to the extent that they lessen undesirable characteristic of RP data by mainly using tradeoffs among variables from hypothetical experiments. Respondent SP Data SP Equilibrium SP Tradeoffs RP Data RP Equilibrium RP Tradeoffs Respondent Figure 2.4 Data Enrichment Paradigm Source: Stated Choice Methods: Analysis and Applications by Louviere et al. (2000) 35 CHAPTER 3 STATEMENT OF RESEARCH The overall goals of this thesis effort were (i) to improve freight transportation modeling by making it more realistic as well as enabling it to represent firmlevel logistical operations aspects and (ii) to investigate how to properly apply disaggregate discrete choice models in the context of freight transport mode and route selection. 3.1 RESEARCH OBJECTIVES According to the research goals mentioned above, the objectives of this thesis can be enumerated as follows; Objective 1: To perform a review of literature related to freight transportation modeling, especially focusing on modal split, flow assignment, multimodal network consideration and inclusion of logistical requirements. Objective 2: To identify criteria for commodities’ logistical characteristics (e.g. time value, physical configuration) and requirements (e.g. bulk/containerized, consignment size) and systematically classify commodity groups according to those criteria. Objective 3: To identify potential explanatory variables which should be included in the utility function when shippers/carriers decide on freight transport mode and routes selection and specify a proper structure of disaggregate discrete choice models (e.g. linear in parameter) based on the identified variables. 36 Objective 4: To investigate the changes to the conventional mathematical program for network flow assignment for applying the proposed disaggregate discrete choice model with the identified explanatory variables. Objective 6: To test the proposed discrete choice model and the network assignment algorithm on a simplified multimodal network with multiple OD pairs and commodities. 3.2 RESEARCH SCOPE AND LIMITATION Because of time and resource constraints, the scope of this research is limited as follows; This research considers only mode and route selection models and excludes flow generation and distribution steps by assuming that freight flows between each origindestination pair are given in the model. The research considers only three modes of transport: (highway, railway and waterway). Each mode can have its own capacity, however, only travel time on the highway network is considered to be affected by congestion. Many commodity groups with different utility functions and different origindestination pairs can be assigned onto the network simultaneously. The multimodal network used to test the developed algorithms is simplified from the real network with link performance data (e.g. travel cost, travel time, travel time variation and so forth). Because parameter estimation is outside the research scope, the parameters of discrete choice model are assumed in this experiment. The mathematical model for network flow assignment considered in this research is stochastic user equilibrium model. 37 3.3 RESEARCH CONTRIBUTION The purpose of this thesis is to contribute to the improvement of strategic freight transportation modeling, especially in mode and route selection tasks. The main contributions can be described as follows; A comprehensive review of literature related to freight transportation modeling, especially, in the aspect of multimodal network and modeling based on logistical requirements and capacity/congestion effect. The improvement of the criteria used to classify commodities by considering logistical characteristics and requirements and the application of clustering approach for the commodity classification. Identification of discrete choice model structure and explanatory variables to be included in the model as well as the algorithm for applying the proposed discrete choice model for stochastic congested network flow assignment. 38 CHAPTER 4 RESEARCH APPROACH To complete the objectives described in Chapter 3, this chapter briefly explains the tasks that must be accomplished. Four main tasks were carried out as part of this research; Identification of criteria for classification of commodity groups based on commodity characteristics and logistical requirements and creation of the commodity groups based on the criteria developed and mathematical models based on clustering approach. Identification of disaggregate discrete choice model structure and explanatory variables to be included in the model. Investigation of the impacts and changes due to the proposed discrete choice model to the conventional mathematical program for stochastic user equilibrium assignment and development of an algorithm for applying the proposed discrete choice model to stochastic congested network flow assignment Proofofconcept implementation of the assignment algorithm on a simplified multimodal network with the assumed models’ parameters. The flow chart of all tasks in the thesis are depicted in Figure 4.1 39 4.1 EXPLANATION OF THE RESEARCH TASKS 4.1.1 Identification of Logistical Requirements and Classification of Commodities Groups Based on the fact that each commodity has its own logistical characteristics and requirements, decisionmaking for freight transport mode and route selection are affected by these issues. For example, perishable goods and computers have high value of time (e.g. their value reduces significantly with the time.) while sand and gravel do not. Another example is that bulk and containerized commodities have different preference and, therefore, requirements for transportation means. Therefore, in order to incorporate Identify logistical criteria and classify the commodities into groups Specification of the discrete choice model’s structure and its explanatory variables Implement the proposed discrete choice model and algorithm on a simplified multimodal network Figure 4.1 Flow Chart of the Thesis Tasks Investigation of the impacts to the conventional mathematical program for network flow assignment and proposal of an algorithm for stochastic network flow assignment 40 this idea into freight transportation modeling, commodity groups should be classified according to their logistical characteristics and requirements. In this section, logistical characteristics and transportation requirements are identified. Then, Standard Classification of Transported Goods (SCTG) developed by the U.S. Department of Transportation, U. S. Bureau of the Census, Statistics Canada, and Transport Canada, which is widely used for freight transport modeling in the U.S. is reviewed and reclassified by using the logistical characteristics and requirements as the criteria. First, logistical characteristics and requirements are reviewed based on available surveys and studies. Then characteristics and requirements important to mode and route selection processes are selected and used as criteria for the classification of commodities. The 4 digit level commodity groups (283 groups) classified by STCG are used as the basis for reclassification. Clustering techniques aiming to maximize relations among commodities in the same group or mathematical programming approach aiming to minimize dissimilarity of commodities in the same group can be applied for systematic classification of commodity groups. 4.1.2 Specification of the Model's Structure and Its Explanatory Variables Basically, each commodity group classified in section 4.1 should have its own set of model parameters based on its own logistical characteristics and requirements. The potential explanatory variables to be included in the discrete choice model can be considered from two aspects: theories of supply chain and logistics and practical perspectives from shippers and carriers interviews. On one hand, from the theoretical aspect, supply chain and logistics theories and studies as well as operational models at the firm level are reviewed. Previous interviews of shippers and carriers to determine 41 influential factors in freight transport mode and route selection are also studied to gain practical insights. Based on the theoretical and practical perspectives, the important explanatory variables to be included in the discrete choice model are identified. The relationship among these explanatory variables are also identified and used for specifying the structure of the discrete choice model. For the sake of simplicity, the discrete choice model structure is assumed to be linear in its parameters (e.g. U ( 1 2 3 L q ,q ,q ,...,q ) = l l q q ... q 0 1 1 2 2 + + + + ) 4.1.3 Proposal of An Algorithm for Network Flow Assignment Based on the Proposed Discrete Choice Model According to the proposed discrete choice model for freight transport mode and route selection, the impacts and changes to the conventional mathematical program for network flow assignment are investigated. An algorithm for network flow assignment based on the proposed discrete choice model is presented. 4.1.4 Implementation of the Discrete Choice Model and Algorithm on a Simplified Multimodal Network The discrete choice model developed and the assignment algorithm are tested on a simplified multimodal network. The multimodal network comprised of the following basic elements. Freight Analysis Framework Version 2.2 highway network (Federal Highway Administration 2006) Railway network (Federal Railroad Administration and Bureau of Transportation Statistics 2006) 42 Waterway network (Vanderbilt Engineering Center for Transportation Operations and Research and Bureau of Transportation Statistics 2006) Intermodal terminals database (Oak Ridge National Laboratory and Bureau of Transportation Statistics 1998) Each individual network will be integrated via intermodal terminals to create a multimodal network. In the proofofconcept implementation, multiple origindestination (OD) pairs and pair multiple commodities are also considered. 43 CHAPTER 5 CRITERIA AND MODELS TO CLASSIFY COMMODITIES INTO LOGISTICAL FAMILIES Conventionally, the classification of commodity groups for freight transport modeling is based on industrial sectors, for example, live animals and fish (SCTG 01) and alcoholic beverages (SCTG 08) (See Appendix A 2.2). However, commodities in the same industrial sector may have different characteristics and logistical requirements. For instance, fresh food and prepared food, both of them are in the food industrial sector and classified into the same SCTG commodity group. However, their characteristics and transportation requirements are very different. Fresh food needs rapid transport mode because it is perishable while prepared food does not have such a requirement. Based on this idea, to increase the accuracy and forecasting capability of freight transport models, the commodity classification should be reconsidered based on their commodity characteristics and logistical requirements. 5.1 THE CONCEPT OF LOGISTICAL FAMILIES Many studies indicated that product characteristics affect the choice of transportation modes and routes of shippers and carriers. Abkowitz et al. (1992) studied 44 the criteria for designating hazardous materials highway routes and found criteria such as accident likelihood and population exposure/risk which are different from routing criteria of other commodities. Pedersen and Gray (1998) indicated in their study that the use of transport modes is clearly affected by the type of product transported. Jiang et al. (1999) described freight demand characteristics as a function of a firm’s characteristics, goods’ physical attributes, and the spatial and flow characteristics of shipments. Cullinane and Toy (2000) also proposed that characteristics of transported goods such as value/weight ratio and density are an important factor affecting mode/route choice. Jose HolguinVeras (2002) studied the commercial vehicle choice process and found that the shipmentsize selection is related to types of commodities. Tatineni and Demetsky (2005) also indicated that value of the commodity, density of a commodity and shelf life of goods are important variables influencing transportation decision process of a firm. The concept of commodity classification based on logistical requirements and characteristics is quite new in the context of freight transport modeling. From the literature review, there are only five studies incorporating this concept into their research. 5.1.1 Strategic Model for Integrated Logistics and Evaluations (SMILE) – the logistical requirements and characteristics used in the model are shown in Table 5.1. Table 5.1: Logistical Characteristics and Requirements in SMILE Model 1) Value/density ratio ($/kg/m3) 2) Packaging density (piece or weight per m3) 3) Value of time (low, medium, high) 4) Delivery strategy (Use or not use DC) 5) Shipment size (small, medium, large) 6) Demand frequency (frequent/ not frequent) 7) Bulk/general cargo 8) Lead time value (low, medium, high) 9) Weight/Volume ratio (kg/m3) In the model, three main characteristics e.g. value/density ratio, packaging density, volume/weight are used to classify commodities into logistical families as shown in Figure 5.1. Figure 5.1 Methodology for Commodity Classification in (Source: Based on Strategic Model for Integrated Logistics and Evaluations by Tavasszy et al., 1998) According to the logistical requirements and characteristics and classification methodology used in the SMILE model, 5.1.2 Swedish National Model System for Freight Transport (SAMGODS) SAMGODS, the logistical density (heavy or light) and value per weight unit (low or high) with other product characteristics such as dry/liquid and consumption/intermediate goods requirements and characteristics Table 5.2. Then commodities are classified into these 12 logistical families ba logistical characteristics and requirements. 45 the SMILE Model 542 types of products are sorted into 50 logistical f characteristics and requirements are bulk/general cargo, are used to develop 12 different groups as illustrated in . families. – In goods. The logistical based on their 46 Freight flow NST/R 2digits Handling category Cereals and agricultural products 00 01 04 05 06 09 17 18 General cargo Consumer food 01 11 12 13 16 Unitised Conditioned food 03 14 Unitised Solid fuels and ore 21 22 23 41 45 46 Solid bulk Petroleum products 31 33 34 Liquid bulk Metal products 51 52 53 54 55 56 General cargo Cement and manufact. build. mat. 64 69 Unitised Crude building materials 61 62 63 65 Solid bulk Basic chemicals 81 83 Solid bulk Fertiliser, plastic and other chem. 71 72 82 84 89 General cargo Large machinery 91 92 939 General cargo Small machinery 931 Unitised Misc. manufacturer articles 94 95 96 97 99 Unitised Table 5.2: Logistical Families in SAMGODS Model 1) Dry and heavy bulk goods, low value 2) Liquid and voluminous bulk goods, low value 3) Investment goods, durable consumer goods, high value 4) Heavy intermediate and consumption goods, low value 5) Lightweight consumer goods with high value 6) Lightweight intermediate and consumer goods with low value 7) Containerized bulk goods 8) Iron ore from Northern Sweden 9) High value container goods 10) Low value container goods 11) Transit goods 12) Air freight 5.1.3 SCENES European Transport Scenarios – Four main handling requirements are considered in this study: unitized, solid bulk, liquid bulk and general cargo. Commodities from Standard goods classification for transport statistics (NST/R) are grouped into 13 logistical families based on the handling requirements as shown in Table 5.3. Table 5.3: Logistical Families in SAMGODS Model 47 5.1.4 Strategic European Multimodal Modelling (STEMM) – The logistical characteristics and requirements used to classify commodities into logistical families are shown in Table 5.4. Commodities from Standard International Trade Classification (SITC) are sorted based on the criteria into 12 logistical families. Table 5.4: Logistical Characteristics and Requirements in STEMM 1) Price (low, medium, high, very high) in EURO/tonne 2) Delivery size (small, medium, large) 3) Density (low, medium, high) in tonne/m3 4) Type of goods (bulk, chemicals, parcelled) 5) Temperature control (yes, no) 6) Risk of damage (low, medium, high) 7) Level of service (low, medium, high). 5.1.5 Relations between demand for freight transport and industrial effects (REDEFINE) – Three main logistical characteristics and requirements are considered in this project: bulk/general cargo, density (high or low) and value (high or low). 14 main commodities are grouped into logistical families according to the criteria. The commodities are agricultural products, beverages and food, wood and paper, building materials, textiles and clothes, other crude minerals, chemicals and fertilizers, petrol and petroleum products, Cola and Coke, metals, machinery, transport equipment, other manufactured articles and miscellaneous articles. The logistical families are shown in Table 5.5. 48 Table 5.5: Logistical Families in REDEFINE Logistical cluster Type of goods Density (ton/m 3 ) Value (Euro/m 3 ) 1 Bulk High Low 2 Bulk Low Low 3 General cargo High High 4 General cargo High Low 5 General cargo Low High 6 General cargo Low Low High density > = 1 (ton/m 3 ) Low density < = 1 (ton/m 3 ) High value > = 5 (Euro/m 3 ) Low value > = 5 (Euro/m 3 ) Based on the review of published research and other related studies, the important logistical characteristics and requirements identified for use in this research are shown in Table 5.6 Table 5.6: Main Logistical Characteristics and Requirements in REDEFINE No. Logistical requirements 1 Value/density ratio ($/kg/m3) 2 Value of time (low, medium, high) 3 Bulk 5 Unitized 6 Weight/Volume ratio (kg/m3) 7 Dry 8 Liquid 9 Temperature control (yes, no) 10 Risk of damage (low, medium, high) 11 Harzardous material (yes, no) 12 Live (yes,no) 13 Holding cost ($/kg/m3 or piece/day) 14 Transshipment cost ($/kg/m3 or piece) It can be seen that the methodologies applied to classify commodities into logistical families is quite subjective and somewhat unsystematic. The classification highly depends on the opinions of researchers in each study. A more systematic method could be applied and results of classification improved. One possible way is to formulate the problem as a mathematical model similar to those in network clustering problems. 49 Clustering can be defined as the process of grouping objects into sets called clusters, so that each cluster consists of elements that are similar in some way. The similarity/dissimilarity criterion can be defined in several different ways, depending on the specific application and the objectives that the clustering aims to achieve. For example, in distancebased clustering two or more elements belong to the same cluster if they are close with respect to a given distance metric. On the other hand, in conceptual clustering, which can be traced back to Aristotle and his work on classifying plants and animals, the similarity of elements is based on descriptive concepts (Balasundaram and Butenko, 2006). 5.2 DISSIMILARITY COEFFICIENTS FOR COMMODITY CLASSIFICATION According to the concept of network clustering, the first task for applying to commodity classification is to define the similarity/dissimilarity of commodities. In this thesis, the dissimilarity coefficient of each pair of commodities is defined based on the weighted Minkowski metric as follows; k ki kj n k 1 ij k Max A A  D W − = = where Dij = Dissimilarity coefficient between commodities i and j Wk = Weight of importance of characteristic and logistical requirement k Aki = Characteristic and logistical requirement k of commodity i 50 Maxk = Maximum value of characteristic and logistical requirement k The details of logistical characteristics and requirements, their values and weights are illustrated in Table 5.7. Table 5.7: Logistical Requirements, Values and Weights to Calculate Dissimilarity Coefficient No. Logistical requirements Value Description Weight 1 Value/density ratio ($/kg/m3) 14 low, med, high 3 2 Value of time (low, medium, high) 13 low, med, high 3 3 Bulk 0,1 Yes/No 3 5 Unitized 0,1 Yes/No 3 6 Weight/Volume ratio (kg/m3) 13 low, med, high 1 7 Dry 0,1 Yes/No 1 8 Liquid 0,1 Yes/No 1 9 Temperature control (yes, no) 0,1 Yes/No 3 10 Risk of damage (low, medium, high) 13 low, med, high 1 11 Harzardous material (yes, no) 0,1 Yes/No 2 12 Live (yes,no) 0,1 Yes/No 3 13 Holding cost ($/kg/m3 or piece/day) 13 low, med, high 3 14 Transshipment cost ($/kg/m3 or piece) 13 low, med, high 3 5.3 OPTIMIZATION MODELS FOR CLASSIFICATION Clustering models can be classified by the constraints on relations between clusters and the objective function used to achieve the goal of clustering (see Balasundaram and Butenko, 2006). One can formulate two types of optimization problems according to a defined measure of cluster cohesiveness as follows; Type I: Minimize the number of clusters while ensuring that every cluster formed has cohesiveness over a prescribed threshold. Type II: Maximize the cohesiveness of each cluster formed while ensuring that the number of clusters that result is under a prescribed number K. The cohesiveness of a cluster in this context is measured using pairwise dissimilarity coefficients of commodities which are defined in section 5.2. The 51 optimization model to classify commodities into logistical families can be formulated in two ways as follows; Type I: Minimize the number of logistical families while limiting the maximum dissimilarity coefficient of a pair of commodities in the same logistical family. Type II: Minimize the cumulative/total dissimilarity coefficient among commodities in the same logistical family according to the specified number of logistical families. The detail of each optimization model are described as follows; Type I Objective function: = n k 1 Minimize yk Subject to x 1 for i 1, 2,..., n n k 1 ik = = = x y for i 1, 2,..., n k 1, 2,..., n ik k £ = = D z D for i 1,2,...,n, j 1,2,...,n and k 1,2,...,n ij ijk £ = = = z x x  1 for i 1,2,..., n, j 1,2,..., n and k 1,2,..., n ijk ik jk ³ + = = = z {0, 1}, y {0,1} and x {0, 1} for i 1,2,...,n, j 1,2,...,n and k 1,2,...,n ijk j ij Î Î Î = = = Where Dij = Dissimilarity coefficient between commodities i and j D = Maximum value of dissimilarity coefficient between commodities i and j to be allowed in the same logistical family xik = 1 if commodity i is assigned to family k; 0 otherwise yk = Logistical family k (if family k is assigned based on commodity k, yk = 1, else yk = 0) 52 zijk = Relationship among product i, product j and logistical family k (if product i and product j are in the same logistical family k, zijk = 1, else zijk = 0 Type II: Objective function: ij n i 1 n j 1 Minimize Dijx = = Subject to x 1 for i 1, 2,..., n n j 1 ij = = = y p n j 1 j = = x y for i 1, 2,..., n j 1, 2,..., n ij j £ = = xij Î {0, 1} for i = 1, 2,..., n j = 1, 2,..., n y {0, 1} for j 1, 2,..., n j Î = Where Dij = Dissimilarity coefficient between commodities i and j xij = 1 if commodity i is assigned to family j; 0 otherwise yj = 1 if family j is formed based on commodity j; 0 otherwise p = the number of logistical families (p is identified by users) n = total number of commodities to be classified Both integer programming formulations are similar to those of the pmedian problem (see Balasundaram and Butenko, 2006) and NPhard in general (see Gonzalez, 1985). However, in the case of commodity classification, the maximum number of commodities to be classified is less than 600 which is efficiently solvable with a good PC 53 and a commercial IP solver. In this research, classification of 49 commodities are formulated based on the developed models and solved by the optimization software XpressMP (http://www.dashoptimization.com/home//products/products_overview.html) on a computer server with a 3056 Mhz processor and 4 GB RAM. For the Type I formulation, it took less than 3 minutes to solve to the optimality, however, for the Type II formulation, it took more than 15 hours before reaching the optimal solution. The main reason for much longer computational time for the Type II formulation is because of the increased number of constraints when the number of commodity increases. 5.4 EXPERIMENTS AND RESULTS OF THE MODELS Both of the proposed approaches are applied to the U.S. commodity classification system, Standard Classification of Transported Commodity (SCTG) 4digit which has 283 groups of commodities. First, the classification of 8 commodities is used as a test case for these approaches. The details of logistical characteristics and requirements for all 8 commodities are illustrated in Appendix A3.1. The results are as follows; Results from Approach 1: Grouping (p = 3) Commodity_Index Description Group 1 1 Live animals and live fish Group 2 2, 5 Fresh or chilled vegetables except potatoes (Irish potatoes), Fresh, chilled, or frozen, except poultry Group 3 3,4,6,7,8 Dried vegetables, Dried fruit, Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered, Gasoline, Fuel oils 54 Grouping (p = 4) Commodity_Index Description Group 1 1 Live animals and live fish Group 2 2, 5 Fresh or chilled vegetables except potatoes (Irish potatoes), Fresh, chilled, or frozen, except poultry Group 3 3,4,6 Dried vegetables, Dried fruit, Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered, Group 4 7,8 Gasoline, Fuel oils Results from Approach 2: Product family Commodities Details 1 2 Fresh or chilled vegetables except potatoes (Irish potatoes) 5 Fresh, chilled, or frozen, except poultry 2 3 Dried vegetables 4 Dried fruit 6 Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered 3 1 Live animals and live fish 4 7 Gasoline 8 Fuel oils Based on the results of the commodity classification shown above, it can be seen that the commodities which were classified into different SCTG groups e.g. fresh/chilled vegetables and fresh/chilled meats in SCTG are now grouped together by these classification methods because they have similar characteristics and logistical requirements. The clustering approach was also applied to a larger set of commodities (49 commodities) in Table 5.8. The details of logistical characteristics and requirements for (D = 2) 55 all 49 commodities are shown in Appendix A3.2. The results of the experiments are as follows; Table 5.8: Commodity Index and Description Commodity Index Description Commodity Index Description 1 Live animals and live fish 25 Silica sands and quartz sands, for uses other than construction, and other sands 2 Wheat 26 Limestone and chalk (calcium carbonate) 3 Corn except sweet 27 Gravel and crushed stone except dolomite, slate, and limestone 4 Fresh or chilled vegetables except potatoes (Irish potatoes) 28 Nonagglomerated bituminous coal 5 Dried vegetables 29 Agglomerated coal 6 Fresh or chilled edible fruit except citrus 30 Fuel oils 7 Dried fruit 31 Pharmaceutical products 8 Freshcut flowers 32 Paints and varnishes 9 Unmanufactured tobacco 33 Soap, organic surfaceactive agents, cleaning preparations, polishes and creams, and scouring preparations 10 Cereal straw or husks and forage products 34 Sparkignition reciprocating internalcombustion engines for motor vehicles, of a cylinder capacity exceeding 1000 cc 11 Dog or cat food put up for retail sale 35 Parts of internalcombustion piston engines 12 Fresh, chilled, or frozen, except poultry 36 Pumps for liquids 13 Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered 37 Airconditioning equipment 14 Wheat flour, groats, and meal 38 Refrigerating or freezing equipment 15 Baked snack foods 39 Electric motors, generators, generating sets, and rotary converters 16 Frozen baked products 40 Electric or electronic transformers, static converters including rectifiers, and inductors 17 Milk and cream 41 Telephone or telegraph switching apparatus except parts 18 Ice cream or ice milk and their novelties, water ices, and sherbets 42 Electronic entertainment products except parts 19 Frozen vegetables and vegetable preparations 43 Computer equipment 20 Processed or prepared vegetables except frozen, dried, or milled 44 Office equipment 21 Coffee, tea, and spices, except unprocessed coffee and unfermented tea 45 Prepared unrecorded media for audio, video, computer, or other uses 22 Sweetened or flavoured water 46 Prerecorded media 23 Cigarettes 47 Electronic components 24 Silica sands and quartz sands, for construction use 48 Photographic cameras, image projectors, enlargers and reducers, projection screens, negatoscopes, and apparatus and equipment for film developing 49 Industrial processcontrol instruments 56 Results from approach 1: p = 5 Group Commodity Index 1 4,6,8,12,16,17,18,19 2 1,5,7,11,13,15,20,21,22,23,31,32 3 33,2,3,9,10,14,30 4 24,25,26,27,28,29 5 34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49 p = 8 Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 5,7,11,13,15,20,21,22,23,31 4 2,3,9,10,14 5 24,25,26,27,28,29 6 30 7 32,33 8 33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49 p = 11 Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 2,9,10,14 4 22,32,33 5 31 6 41,42,43,44,49 7 24,25,26,27,28,29 8 34,35,36,37,38,39,40,45,46,47,48 9 30 10 3 11 5,7,11,13,15,20,21,23 57 Results from approach 2: Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 2,9,10,14 4 22,32,33 5 31 6 41,42,43,44,49 7 24,25,26,27,28,29 8 34,35,36,37,38,39,40,45,46,47,48 9 30 10 3 11 5,7,11,13,15,20,21,23 According to the results of the classification of 49 commodities, commodities with similar characteristics and logistical requirements are grouped together. For example, in case that p = 11 for Type I formulation and D = 2 for Type II formulation, wheat, unmanufactured tobacco, cereal straw and wheat flour are grouped together. (D = 2) 58 CHAPTER 6 IDENTIFICATION OF EXPLANATORY VARIABLES AND THE STRUCTURE OF DISCRETE CHOICE MODEL In this chapter, the main purpose is to examine how logistics and supply chain management practices at the firmlevel relate to freight transport mode/route selection decisionmaking and how this concept can be incorporated into a macrolevel freight movement model. The consideration can be explored from two aspects: theories of supply chain and logistics perspective and practical perspective from shippers and carriers interview. At the end, based on both perspectives, the influential variables which should be included in the transport mode/route selection model and the structure of the model are identified. 6.1 VARIABLES BASED ON SUPPLY CHAIN AND LOGISTICS THEORETICAL PERSPECTIVE The very first thing that can be observed is the high level of interrelation between transportation decision and logistics/supply chain management at a firm level. Tatineni and Demetsky (2005) indicated that individual firms take transportation decisions as a part of the larger process of optimizing the total supply chain performance or minimizing the total supply chain costs rather than minimizing only the transportation costs. 59 SimchiLevi et al. (2003) described the relation among shipment sizes, inventory holding costs and ordering costs including transportation costs as exhibited in Figure 6.1. Figure 6.1 Relations among Shipment Size, Inventory Costs and Transport Costs Source: Designing and Managing the Supply Chain by SimchiLevi et al. (2003) From Figure 6.1, it can be obviously seen that transport choices are closely related to shipment size and inventory holding cost. More precisely, the figure shows the tradeoff between inventory cost and transport cost as shipment size varies. If one chooses to transport goods with a small shipment size, one will pay more for direct transport cost but less for inventory cost and vice versa. According to this close relation between transportation and shipmentsize decisions, many researchers proposed that the decision for transport mode and shipment size should be considered simultaneously in freight transport model. Baumol and Vinod (1970) applied inventory theory for freight transport demand model. According to their model, total cost of transport and inventory includes direct shipping costs, intransit carrying costs, ordering costs, inventory carrying costs and costs of safety stock. In the model, unit Q* Order Quantity Q Total Cost Holding Cost Ordering Cost 60 shipping cost, transit time and unit inventory carrying cost are characterization of a mode of transport. Ordering costs and inventory carrying costs are the function of shipment frequency or shipment size. Therefore, their model explicitly included the tradeoff between transport mode and shipment size decisions. Das (1974) argued that the selection of transportation services in a firm should not consider only the direct transport cost but should consider the total cost of logistics including total direct shipping cost, intransit carrying cost and the cost of consignee’s inventory operations. Constable and Whybark (1978) developed a mathematical model to minimize total logistics costs at a firm level including transportation cost, intransit inventory cost, order cost, inventory carrying cost and back order cost. The decision variables in their model can be divided into two groups: characteristics of transport modes (e.g. transport cost and transit time) and inventory variables (e.g. reorder point and order quantity). Chiang et al. (1981) proposed freight demand model which can simultaneously forecast transport mode and shipment size. The alternatives in their model are the combination of transport modes and shipments. The utility is the function of total logistics costs comprising of transport rate, capital carrying cost, intransit carrying cost, order cost and loss of value during transit and storage cost. Abdelwahab and Sargious (1992) proposed twostage conditional discrete choice model for shipment size and transport mode. First, shipment sizes of rail and road transport mode are calculated and then the calculated shipment sizes of these two modes are used as the explanatory variables to calculate transport mode choice. The attributes used in the models as explanatory variables are commodity attributes (e.g. value and density), modal attributes (e.g. freight charges, reliability and transit time) and other attributes such as regional variables, yearly demand. 61 Ostlund et al. (2002) proposed the discrete choice model for transport mode selection. In the model, total logistics costs include warehousing costs, safetystock costs at warehouse, cyclestock costs at warehouse, intransit inventory costs, trunking and delivery costs. Transport related costs (e.g. trunking and delivery costs) are the function of shipment size and this makes transport cost function nonlinear. Therefore, they proposed a linear approximation of the transport cost function by dividing shipment size into three classes: less than truck load (LTL), full truck load (FTL) and train wagon load (WL). Federal Railroad Administration (2005) developed the intermodal transportation model to study utilization between highway and railway. It was indicated that the choice of shipment sizes can affect the choice of transport modes. De Jong and BenAkiva (2007) proposed a microsimulation model for shipment size and transport chain selection. The optimal shipment size, in this research, is the one which minimizes the total logistics cost including order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs. The optimal shipment size is then used to calculate the transport mode which minimizes the total logistics cost. An important assumption for this method is that transport costs do not matter in the determination of shipment size. Based on all of the above mentioned studies, it is obvious that when a firm makes decisions for transportation services, direct transportation costs are not the only factors considered. Logistical cost components such as ordering costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs are 62 considered simultaneously. Therefore, in the context of freight modeling, these logistics costs should be incorporated into the model to better represent the decisionmaking processes of shippers and carriers. Another noticeable point is that shipment size selection largely affects the transport mode selection. 6.2 VARIABLES BASED ON PRACTICAL SHIPPERS AND CARRIERS PERSPECTIVES In section 6.1, the theoretical aspects of supply chain and logistics management have been explored. In this section, we will examine potential variables which can influence transport mode and route decisions from the practical perspectives of shippers and carriers according to conducted surveys and interviews. The details are as follows. Bardi (1973) examined the carrier selection criteria for household goods manufacturers. The five most important considerations are as follows; Reliability (meeting estimated pickup and delivery dates) Security (frequency of damage, ease of claim settlement, extent of damage) User satisfaction (courtesy of carrier employee, employee complaints, carrier reputation) Availability (carrier representative, nationwide operating authority) Transit time and cost Greeno et al. (1977) investigated the transport attributes affecting transport mode selection comprising of rail freight, rail express, private fleet, steamship, truck forwarder and common carrier. The investigated list of attributes is as follows;  Ontime delivery  Time in transit  Expensiveness  Frequency of service  Tracing time  Completeness of service 63  Promptness of claim and settlement  Competence of staff  Equipment availability  Flexibility of service  Simplicity of dealing  Intermodal flexibility  Care with shipments  Innovativeness  Quality of personal According to the interview and questionnaire of 80 firms, it was found that six attributes were the most important and almost totally explain the decisions. These six attributes are time in transit, inexpensiveness, frequency of service, tracing time, flexibility of service and care with shipments. McGinnis (1980) applied the factor analytic approach to examine factors influencing shippers’ choice of transportation modes. The results ranked in order of importance can be shown as follows; Speed and reliability Freight rates Loss and damage External market influences Inventories Market competitiveness Company policy and customer influences McGinnis (1989) investigated the results of 11 studies related to transportation mode choice. He concluded that the factors affecting transport mode selection could be classified as follows; Freight rates – costs, charges, rates Reliability – reliability, delivery time 64 Transit time – timeintransit, speed, delivery time Over, short and damaged  loss, damage, claim processing and tracing Shipper market considerations – customer service, user satisfaction, market competitiveness, market influence Carrier considerations – availability, capability, reputation, special equipment Product characteristics – product perishability, packaging requirements, new products Jeffs and Hills (1990) studied the determinants of transport mode choice of shippers in the paper, printing and publishing industry in Britain. They found that the determinants are reliability, control over dispatch and delivery time, avoidance of damage to goods, security of product in transit, transit time, ready to transport when required, length of haul and size of consignment. Abkowitz et al. (1992) studied the criteria for highway route choices of hazardous materials transportation. They proposed that, at least five criteria should be considered for route selection to make the decision safe for public and efficient simultaneously. The criteria are shipment, distance, travel time, accident likelihood, and population exposure. Indeed, the multiplication of accident likelihood and population exposure is the classic definition of risk in hazardous material transportation. Matear and Gray (1993) studied the factors influencing freight service choice for shippers and freight suppliers in Irish sea freight market. Of thirty investigated factors, the most important factors included fast response to problems, avoidance of loss or damage, ontime collection and delivery, value for money price, good relationship with carrier, ability to perform unanticipated urgent deliveries, short transit time, low price and ability to handle shipments with special requirements. Murphy et al. (1997) studied the transport mode selection criteria of shippers and carriers and compared the difference. The results are shown in Table 6.1. 65 Table 6.1: Mean Score of Factors Influencing Transport Mode Selection of Shippers and Carriers Shipper Carrier Reliability 1.22 1.28 Equipment availability 1.36 1.79 Transit time 1.45 1.83 Pickup and delivery service 1.60 1.93 Financial stability 1.63 2.07 Operating persennel 1.66 1.81 Loss and damage 1.69 2.11 Rates 1.71 2.57 Service frequency 1.83 2.09 Scheduling flexibility 1.85 1.96 Expediting 1.88 2.07 Rate changes 1.88 2.58 Service changes 1.89 2.05 Tracing 1.91 2.63 Linehaul services 1.98 2.33 Claims 2.05 2.72 Carrier salesmanship 2.68 2.89 Special equipment 3.05 2.51 Mean score Factor Source: Carrier Selection: Do Shippers and Carriers Agree, Or not by Murphy et al. (1997) As shown in Table 6.1, the relative importance of influential attributes is quite the same between shippers and carriers. The most important factors are reliability, equipment availability, operating personnel, transit time, pickup and delivery service and financial stability. Pedersen and Gray (1998) studied the transport selection criteria of shippers in Norway. The investigated determinants included Timing factors – reliability in collection and delivery time, high transport frequency, short transit time and directness of the transport route Pricing factors – freight rate, difference between actual and estimated costs, special offer/discount, packing charges Security factors – damage/loss frequency, control over delivery time, ability to 66 monitor the goods in transit, knowledge of port/labor Service factors – coordination and cooperation with carriers, flexibility, ability to handle urgent deliveries and ability to handle special consignments They concluded that, for Norwegian shippers, transport price factors are the most important ones. However, the influential factors highly depend on product characteristics. For example, the exporters with high valuetoweight ratio consider the timing factors more important than pricing factors. Jiang et al. (1999) studied the demand characteristics affecting transport mode selection from revealed preference data in France. They classified demand characteristics into three types as follows; A firm’s characteristics include type of firm (for example, factories, shopping centers, or warehouses), the firm’s structure (small, nationwide or worldwide) and the firm’s location (for example, the accessibility to rail branch lines and highways). A firm’s own transportation facilities closely relate to its transportation demand and are also an important factor in its modal choice. In addition, a firm’s information system strongly influences its logistic practices and plays an increasingly important role in its transportation decisions. Another demand characteristic is attributes of the goods to be transported, such as type of product, weight, value, and packaging. Packaging is generally either parcels and pallets, containers, and cases. The last demand characteristic is spatial distribution and physical flow including frequency, distance and origin/destination of shipments. It was found that transportation distance, company size and type, information system, accessibility to transport infrastructure from origin and destination, shipment packaging, 67 truck ownership, shipment size are critical determinants for transport mode decision. Cullinane and Toy (2000) reviewed 75 articles related to freight transport mode/route selection criteria and applied the content analysis to identify critical attributes that have been most mentioned in the literature. The list of the attributes mentioned in the reviewed literature is shown in Table 6.2. According to the results of the content analysis, the most five important factors are cost/price/rate, speed, transit time reliability, characteristics of goods and service (unspecified). Norojono and Young (2001) investigate the shippers’ perception of rail freight services in Indonesia. In this research, the investigated attributes are as follows; Transport charge Delivery time Service quality represented by delivery time reliability measured by the probability of being late due to irregular service, safety with respect to the number of cargoes having loss or damage, distance to rail terminal, and train type whether it is freight train or part of passenger train Flexibility of the service represented by frequency measured by the number of service in each day, departure time, responsiveness measured by complaints The results of the study showed that reliability of delivery time and loss/damage are the major influences. In addition, it is reasonable to define flexibility as the function of frequency and responsiveness. 68 Table 6.2: Category of Attributes and Underlying Terms Category Name Terms covered by category Cost/Price/Rate Cost, price, rate Service (nonspecified) Service (nonspecified) Transit time reliability Transit time reliability Frequency Frequency Distance Distance Speed Speed, transit time, terminal time, transshipment time Flexibility Flexibility, convenient schedule, nonspecific extras, pickup and delivery Infrastructure availability Infrastructure availability, accessibility Capability Capability, service availability, equipment availability, capacity Inventory Inventory Loss/Damage Loss/Damage, claims Characteristics of the goods Type, value, value/weight ratio, volume, weight, density, shipment size Sales per year Sales per year Controllability/Tracability Controllability, tracability Previous experience Nonspecific positive behavior, relationship, image of modes used, stability of firm Source: Identifying Influential Attributes in Freight Route/Mode Choice Decisions: A Content Analysis by Cullinane and Toy (2000) Jose HolguinVeras (2002) studied the choice of shipment size and types of vehicle. It is indicated that shipment size could be adequately modeled as a function of the trip distance, the type of commodity being transported, and the type of economic activity taking place at the origin and destination of the trip. The economic activities at origin and destination, in this case, are retail, wholesale and other activities. Tuna and Silan (2002) investigate the selection criteria for transportation services of liner shippers in Turkey. Table 6.3 shows the relative importance of each criterion. 69 Table 6.3: Relative Importance of Transport Selection Criterion Overall Standard Mean Score Deviation Delivering the cargo without damage 4.84 0.44 Issuing accurate shipping documentation 4.81 0.47 Delivering the cargo at the promised time 4.68 0.63 Dependability in handling problems 4.68 0.53 Informing of changes to schedules 4.62 0.55 Issuing accurate price quotations 4.54 0.77 Responding to complaints quickly 4.54 0.77 Issuing shipping documentation quickly 4.51 0.77 Responding to urgent deliveries quickly 4.49 0.93 Giving clear & correct information about costs 4.43 0.87 Issuing accurate invoices 4.43 0.87 Transit time 4.38 0.86 Willingness of the personnel to help 4.35 0.72 Responding to enquiries promptly 4.35 0.79 Minimum changes to schedules 4.27 0.9 Providing clean& undamaged equipment 4.27 0.87 Expert and knowledgeable personnel 4.19 0.81 Informing whether goods will be transshipped 4.16 0.83 Polite and Respectful personnel 4.11 0.77 Informing about the condition of the cargo 4 0.91 Giving arrival notices on time 4 1.18 Convenient working hours for contact 3.89 0.84 Issuing invoices on time 3.73 1.02 Providing special equipment 3.62 0.92 STATEMENTS Source: Freight Transportation Selection Criteria: An Empirical Investigation of Turkish Liner Shipping by Tuna and Silan (2002) As seen from Table 6.3, ‘Delivering the cargo without damage’, ‘issuing accurate shipping documentation’, ‘delivering the cargo at the promised time’ and ‘dependability in handling problems’ were determined as the most important factors. Mangan et al. (2002) studied the attributes affecting the selection of port and ferry services in Ireland. In their study, there are 14 important attributes affecting the choice of port and ferry services ranked in order of importance as shown in Table 6.4. 70 Table 6.4: Influential Attributes for Port/Ferry Choice (Ranked in Order) 1. Space available when needed on ferry 2. Sailing frequency/convenient sailing times 3. Risk of cancellation/delay 4. Port and ferry on fastest overall route 5. Proximity of ports to origin/destination 6. Cost of ferry service/discounts 7. Speed of getting to/through ports 8. Port/ferry on cheapest overall route 9. Ferry suitable for special cargo 10. Delays due to driving bans, etc. 11. Availability of info on sailing options 12. Facilities for drivers 13. Opportunity for driver rest break 14. Intermodal/connecting transport links Rank of Important Attributes for Port/Ferry Choice Source: Modelling Port/Ferry Choice in RoRo Freight Transportation by Mangan et al. (2002) MorenoQuintero and Watling (2002) studied the route choice behaviors of truck drivers in Mexico and found that variables affecting such a decision were direct travel cost, fines for overloading, enforcement strategies, level of service and congestion. Beuthe et al. (2003) studied attributes influencing transport mode selection by using stated preference survey. The commodities were steel, textile, electronics, chemical, cement, packing and pharmacy. The transport modes included rail, road, waterway, shortsea shipping and intermodal transport. The attributes expected to be influential are as follows; Cost, i.e. outofpocket cost for transport, including loading and unloading; Time, i.e. doortodoor transport time, including loading and unloading; Loss as the % of commercial value lost from damages, stealing and accidents; Frequency of service per week proposed by the carrier or the forwarder; Reliability as the % of deliveries at the scheduled time; Flexibility as the % of times nonprogrammed shipments are executed without undue delay. 71 The results from their analysis showed that the importance of each attribute is different for the different commodities. For instance, time and reliability are important for the textile firm and the producer of electronics, which ship over rather long distances. Reliability, flexibility and losses appear important for the pharmaceutical firm, which seems ready to pay for it. Vannieuwenhuyse et al. (2003) conducted an online survey of shippers and logistics service providers in Belgium to collect data about attributes affecting selection of transport modes. Five most important attributes were identified: transportation cost, reliability, flexibility, transportation time and safety. Punakivi and Hinkka (2006) investigated the selection criteria for transportation modes of shippers in four Finnish industry including electronics, pharmaceutical, machinery and construction. They concluded that the selection criteria for transportation modes are different depending on product characteristics in each industrial sector. The selection criteria for each industry ranked in order of importance are presented in Table 6.5. Table 6.5: Selection Criteria for Transportation Modes Rank of Importance Electronics Pharmaceutical Machinery Construction 1 Quality Speed Price Price 2 Speed Convenience Reliability Scheduling 3 Price Safety Punctuality Punctuality 4 Convenience Fluency Speed Convenience Source: Selection Criteria of Transportation Model: A Case Study in Four Finnish Industry Sectors by Punakivi and Hinkka (2006) Note: quality covers reliability, accuracy and safety. Convenience represents the ability to take special product characteristics into account in operations. Based on the above studies, it is evident that there are some factors besides transportation and logistics costs influencing transport mode and route decisions. However, at the macro level freight demand model, the factors which should be 72 considered are shown in Table 6.6. Table 6.6: Main Influential Factors at Macro Level 1 Transit time 2 Transit time variability 3 Frequency of loss/damage 4 Ease of claim settlement 5 Frequency of service 6 Tracing capability 7 Availability of special handling equipment 8 Flexibility of service 9 Product characteristics 10 Accident likelihood 11 Population exposure 12 Accessibility to transport infrastructure 13 Economic activities at origin/destination 14 Enforcement on highway route Factors Affecting Transport Mode/Route 6.3 RELATIONSHIP AMONG ATTRIBUTES In this section, the relationship among all factors described in section 6.1 and 6.2 are considered. From the review in section 6.1 and 6.2, it can be concluded that freight transportation mode/route decisions can be affected by the factors as follows; Shipment size Total logistics cost including order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs Service quality attributes including transit time, transit time variability, frequency of loss/damage, ease of claim settlement, frequency of service, tracing capability, availability of special handling equipment, flexibility of service, product characteristics, accessibility to transport infrastructure, economic activities at an 73 origin/destination, enforcement on highway route, accident likelihood and population exposure (Note: the last two attributes are for hazardous materials) All of the above influential factors can be included into the model as follows; For shipment size, this factor can be incorporated into the model either as a decision variable or an explanatory variable. However, as exhibited in Figure 6.1, the transport decision is highly dependent on shipment sizes. In addition, based on the previous research experiences by Chiang et al. (1981) and De Jong and BenAkiva (2007), it has been shown that it is reasonable to construct the discrete choice model for transport mode and shipment size simultaneously. Therefore, in this thesis, shipment size will be incorporated into the discrete choice model as a decision variable. Figure 6.2 shows the interaction among all related factors e.g. logistics costs, transport mode/route characteristics, shipment size and other related factors According to Figure 6.2, shipment size interacts with all types of logistics costs e.g. order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs Shipment size can be affected by product characteristics and an activity at origin/destination such as wholesale, retail or manufacturing Transit time and transit variability affect transport costs, capital costs of goods in transit, inventory costs, capital costs of goods in inventory and stockout costs Frequency of loss/damage affects costs of damage during transit 74 Figure 6.2 The Interaction among Related Factors Transport Mode/Route Characteristics  Transit time  Transit time variability  Frequency of service  Ease of claim settlement  Frequency of loss/damage  Tracing capability  Availability of special handling equipment  Flexibility of service  Accident likelihood  Population exposure  Enforcement on highway route Other Factors  Product characteristics  Accessibility to transport infrastructure  Activity at origin/destination Total Logistics Costs  Order costs  Transport costs  Consolidation/distribution costs  Costs of damage during transit  Capital costs of goods in transit  Inventory costs and capital costs of inventory  Stockout costs Shipment Size 75 6.4 IDENTIFICATION OF INFLUENTIAL ATTRIBUTES AND MODEL STRUCTURE This section explains how to include all explanatory variables described in section 6.3 into the discrete choice model. The details are as follows; 6.4.1 Shipment Size – In this thesis, shipment sizes are divided into 3 intervals: 0 ton shipment sizes < 25 tons (e.g. equivalent to 1 truckload), 25 tons shipment sizes < 50 tons (e.g. equivalent to trailer truck) and 50 tons shipment sizes < 80 tons (e.g. more than a trailer truck but with in capacity of rail and ship). The combination of a shipment size and a transportation path forms an alternative in the discrete choice model. 6.4.2 Transit Time –Total transit time of a path from an origin to a destination is the summation of transit time of each link and at each transshipment node in that path as shown in Figure 6.3. For railway/waterway links and transshipment node, they are assumed to be uncapacitated and their transit times are assumed to be normally distributed with a constant mean. However, for a highway link, it is assumed to be capacitated and its transit time is dependent on link’s flow, capacity and free flow travel time in the following forms; = + H ) Q V TT FT 1 ( Where TT is transit time FT is freeflow travel time of a highway link 76 V is traffic flow on that link Q is capacity of the link and are parameter ( is 0.15 and is 4, Bureau of Public Roads, 1964) Figure 6.3 Transit Time of A Path 6.4.3 Transit Time Variability – In this thesis, transit time and transshipment time are assumed to be normally distributed and independent from those of other links. Transit time variability can be measuered by variance of transit time which can be calculated based on coefficient of variation and transit time of a link e.g 2 2 = (C.V. × mean transit time) . Total variance of transit time
Click tabs to swap between content that is broken into logical sections.
Rating  
Title  Tapplication of Disaggregate Discrete Choice Model for Intermodal Stochastic Congested Freight Network Flow Assignment 
Date  20090501 
Author  Sittivijan, Peerapol 
Department  Industrial Engineering & Management 
Document Type  
Full Text Type  Open Access 
Abstract  Nowadays, commercial companies view transportation process as a part of the whole logistic concept. However, most existing freight transport models are still focused only on direct factors such as transport cost and transit time. In this thesis, an alternative way for freight transport modeling which also considers other important factors in the context of supply chain and logistics was explored. The proposed models are expected to increase accuracy and forecasting capability as well as address logistical considerations e.g. shipment size, use of distribution/consolidation centers and inventory aspects within the existing freight transport models. The first part of the thesis developed optimization models based on the clustering concept for classifying commodities into logistical families. The second part of the thesis includes a study of explanatory variables and the structure of discrete choice model for freight transport mode and route selection. In addition, the conventional mathematical model for stochastic user equilibrium assignment and the problems when applying such a model with the proposed utility function were analyzed. A heuristic approach for stochastic user equilibrium with the proposed utility function was developed and tested with a small intermodal network for multi OD pairs and multi commodity assignment. For the first part of the thesis, the optimization models developed were applied to classify commodities into logistical families. The numerical experiments showed that the algorithms were flexible and effective in classifying commodities. For the second part, it was shown that the new utility function incorporating important supply chain and logistics variables made the conventional mathematical program for stochastic user equilibrium not equivalent to the flow pattern at the equilibrium point. As a result, a heuristic algorithm for stochastic user equilibrium assignment was developed and tested with multiple OD pairs/commodities and freight flow assignment was illustrated using a simplified intermodal network. Based on the numerical example, the proposed heuristic algorithm appears to function quite efficiently. 
Note  Thesis 
Rights  © Oklahoma Agricultural and Mechanical Board of Regents 
Transcript  ii TAPPLICATION OF DISAGGREGATE DISCRETE CHOICE MODEL FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT By PEERAPOL SITTIVIJAN Master of Engineering Asian Institute of Technology Pathumthani, Thailand 2001 Submitted to the Faculty of the Graduate College of the Oklahoma State University In partial fulfillment of the requirements for the Degree of MASTER OF SCIENCE May, 2009 ii APPLICATION OF DISAGGREGATE DISCRETE CHOICE MODELS FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT Thesis Approved: Dr. Manjunath Kamath___________ Thesis Adviser Dr. Ricki. G. Ingalls______________ Dr. Tieming Liu_________________ Dr. Balabhaskar Balasundaram_____ Dr. A. Gordon Emslie____________ Dean of the Graduate College iii ACKNOWLEDGEMENTS It is difficult to express my gratitude for the amount of help that my advisor, Prof. Dr. Manjunath Kamath, rendered in every respect, for being patient and for giving me an opportunity to carry out this research work as well as giving various valuable suggestions throughout my staying here as a Master's student. I owe a lot to him for providing advice regarding the course selection, moral and financial support and last but not the least for proof reading this thesis. I would also like to thank the committee members, Dr. Ricki Ingalls and Dr. Tieming Liu as well as Dr. Balabhaskar Balasundaram, for providing valuable information and insights into my research. I would also take this opportunity to thank all my friends in Center for Computer Integrated Manufacturing Enterprises (CCiMe) and Center of Engineering Logistics and Distribution (CELDi) for their support at various stages of my thesis. Finally, I would like to dedicate this thesis to my parents, Mr. Vijan Sittivijan, my father, and Mrs. Luksanee Sittivijan, my mother and all other close relatives for their love and support. I owe a lot to them for constantly motivating me with a lot of advice every time I need. iv LIST OF CONTENTS INTRODUCTION………………………………………………………………….……………1 1.1 TAXONOMY OF FREIGHT TRANSPORTATION MODELS………………………..2 1.2 DISAGGREGATE DISCRETE CHOICE MODELS…………………………………...4 1.3 PROBLEMS STATEMENT AND NEED FOR THE RESEARCH……………….…....5 1.4 OUTLINE OF THE THESIS…………………………………………………………….6 LITERATURE REVIEW……………………………………………………………………….8 2.1 REVIEW OF FREIGHT TRANSPORTATION MODELS IN PRACTICE.....………....8 2.2 TRANSPORTATION MODAL SPLIT AND FLOW ASSIGNMENT………………...15 2.3 MULTIMODAL NETWORK MODELS……………………………………………....23 2.4 CLASSIFICATION OF COMMODITIES BASED ON LOGISTICAL REQUIREMENTS……………………………………………………………………....24 2.5 APPLICATIONS OF DISAGGREGATE DISCRETE CHOICE MODELS IN TRANSPORTATION…………………………………………………………………....26 2.6 REVEALED AND STATED PREFERENCE SURVEY FOR PARAMETER .ESTIMATION…………………………………………………………………………..31 STATEMENT OF RESEARCH…..……………………………………………………………35 3.1 RESEARCH OBJECTIVES……..……………………………………………………....35 3.2 RESEARCH SCOPE AND LIMITATION..…………………………………………….36 3.3 RESEARCH CONTRIBUTION……...………………………………...…………..…....37 v RESEARCH APPROACH………..……………………………………………………………38 4.1 EXPLANATION OF THE RESEARCH TASKS ....…….............…………………......39 CRITERIA AND MODELS TO CLASSIFY COMMODITIES INTO LOGISTICAL FAMILIES…………………….………………………………………………..………………..43 5.1 THE CONCEPT OF LOGISTICAL FAMILIES…………………….………………….43 5.2 DISSIMILARITY COEFFICIENTS OF COMMODITIES……………………….….....49 5.3 OPTIMIZATION MODELS FOR CLASSIFICATION………………………………...50 5.4 EXPERIMENTS AND RESULTS OF THE MODELS……………………………...…53 IDENTIFICATION OF EXPLANATORY VARIABLES AND THE STRUCTURE OF THE DISCRETE CHOICE MODEL……………………...…………………………………..58 6.1 VARIABLES BASED ON SUPPLY CHAIN AND LOGISTICS THEORETICAL PERSPECTIVE..........................................................………………..…….……………58 6.2 VARIABLES BASED ON PRACTICAL SHIPPERS AND CARRIERS PERSPECTIVES .........................................................................................….….....…...62 6.3 RELATIONSHIP AMONG ATTRIBUTES...........................................….…..…..…….72 6.4 IDENTIFICATION OF INFLUENTIAL ATTRIBUTES AND MODEL STRUCTURE......75 6.5 MODIFIED UTILITY FUNCTION FOR PATH OVERLAPPING PROBLEM..……...83 ALGORITHM FOR INTERMODAL STOCHASTIC CONGESTED FREIGHT NETWORK FLOW ASSIGNMENT..........................................................................................85 7.1 HEURISTIC ALGORITHM FOR STOCHASTIC USER EQUILIBRIUM FREIGHT NETWORK FLOW ASSIGNMENT................................................................................85 7.2 NUMERICAL EXPERIMENT.........................................................................................90 SUMMARY AND FUTURE RESEARCH....................................…………………………...107 REFERENCES……………………..……………………………………………..…………....109 vi THEORY, DERIVATION AND PROPERTIES OF DISAGGREGATE DISCRETE CHOICE MODEL..………………..…………………………………………………………..122 A1.1 DISCRETE CHOICE AND RANDOM UTILITY THEORY…………………..…...122 A1.2 CHOICE SET DETERMINATION AND FORMATION…………………………...129 A1.3 INDEPENDENCE FROM IRRELEVANT ALTERNATIVES PROPERTY (IIA)…………………………………….....…………………………....130 A1.4 STATE OF THE ART IN DISAGGREGATE DISCRETE CHOICE MODEL RESEARCH ………………………………………….....…………………………...133 FIRST LEVEL (2DIGIT) STCC AND STCG DEFINITION………………………...……136 THE LOGISTICAL CHARACTERISTICS AND DISSIMILARITY COEFFICIENTS...138 A3.1 LOGISTICAL CHARACGTERISTICS OF 8 COMMODITIES.......…………..…...138 A3.2 LOGISTICAL CHARACGTERISTICS OF 49 COMMODITIES....... ………..…...139 CONVENTIONAL MATHEMATICAL PROGRAM FOR STOCHASTIC USER EQUILIBRIUM ASSIGNMENT WITH TRAVEL TIME UTILITY FUNCTION............143 THE DETAILS OF EXPERIMENT........................................................................................154 A5.1 THE DETAILS OF LINK CHARACTERISTICS...........................…………..….....154 A5.2 PATHLINK INCIDENCE MATRIX..............................................…………...…....155 A5.3 TRANSPORTATION COSTS BY MODES.......................................................…....156 A5.4 CALCULATION OF TOTAL UTILITY OF EACH ALTERNATIVE OF MEAT/SEAFOOD...............................................................................................…....157 A5.5 CALCULATION OF TOTAL UTILITY OF EACH ALTERNATIVE OF NATURAL SANDS.............................................................................................…....160 vii LIST OF FIGURES FIGURE 2.1 LINKS FEEDING TRIPS INTO A NODE..........................................................21 FIGURE 2.2 SIMPLE NETWORK OF WATERWAYS..........................................................23 FIGURE 2.3 CORRESPONDING VIRTUAL NETWORK.....................................................23 FIGURE 2.4 DATA ENRISHMENT PARADIGM...................................................................34 FIGURE 4.1 FLOW CHART OF THE THESIS TASKS.........................................................39 FIGURE 5.1 METHODOLOGY FOR COMMODITY CLASSIFICATION IN THE SMILE MODEL......................................................................................45 FIGURE 6.1 RELATIONS AMONG SHIPMENT SIZE, INVENTORY COST AND TRASPORT COSTS..............................................................................................59 FIGURE 6.2 THE INTERACTION AMONG RELATED FACTORS..................................74 FIGURE 6.3 TRANSIT TIME OF A PATH..............................................................................76 FIGURE 6.4 TRANSIT TIME VARIABILITY OF A PATH..................................................76 FIGURE 7.1 INTERMODAL NETWORKS OF HIGHWAY, RAILWAY AND WATERWAY.........................................................................................................92 FIGURE 7.2 SIMPLIFIED INTERMODAL NETWORKS.....................................................92 FIGURE A1.1 AN EXAMPLE OF NESTED LOGIT MODEL FOR RED BUS BLUE BUS SERVICES..................................................................................132 FIGURE A4.1 EXAMPLE NETWORK FOR STOCHASTIC USER EQUILIBRIUM ASSIGNMENT........................................................148 viii LIST OF TABLES TABLE 2.1 SUMMARY OF THE REVIEW OF FREIGHT MODELS IN EUROPE AND THE U.S.............................................................................................................14 TABLE 2.2 SUMMARY OF FREIGHT TRASPORT MODAL SPLIT MODELS...............17 TABLE 2.3 SUMMARY OF ALGORITHMS USED FOR FLOW ASSIGNMENT.............22 TABLE 2.4 LEVELS OF CLASSIFICATION IN STCC.........................................................25 TABLE 2.5 LEVELS OF CLASSIFICATION IN SCTG.........................................................26 TABLE 2.6 CRITERIA FOR PATH FEASIBILITY................................................................29 TABLE 5.1 LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN SMILE MODEL.................................................................................................44 TABLE 5.2 LOGISTICAL FAMILIES IN SAMGODS MODEL...........................................46 TABLE 5.3 LOGISTICAL FAMILIES IN SAMGODS MODEL...........................................46 TABLE 5.4 LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN STEMM...............................................................................................................47 TABLE 5.5 LOGISTICAL FAMILIES IN REDEFINE...........................................................48 TABLE 5.6 MAIN LOGISTICAL CHARACTERISTICS AND REQUIREMENTS IN REDEFINE.......................................................................48 TABLE 5.7 LOGISTICAL REQUIREMENTS, VALUES AND WEIGHTS TO CALCULATE DISSIMILARITY COEFFICIENT.............................................50 TABLE 5.8 COMMODITY INDEX AND DESCRIPTION.....................................................55 TABLE 6.1 MAIN SCORE OF FACTORS INFLUENCING TRANSPORT MODE SELECTION OF SHIPPERS AND CARRIERS.................................................65 ix TABLE 6.2 CATEGORY OF ATTRIBUTES AND UNDERLYING TERMS......................68 TABLE 6.3 RELATIVE IMPORTANCE OF TRANSPORT SELECTION CRITERION..69 TABLE 6.4 INFLUENTIAL ATTRIBUTES FOR PORT/FERRY CHOICE........................70 TABLE 6.5 SELECTION CRITERIA FOR TRANSPORTATION MODES........................71 TABLE 6.6 MAIN INFLUENTIAL FACTORS AT MACRO LEVEL..................................72 TABLE 6.7 INFLUENTIAL FACTORS INCLUDED AT THE MACRO LEVEL...............80 TABLE 7.1 DETAILS OF COMMODITIES TRANPORTED AMONG OD PAIRS...........91 TABLE 7.2 DETAILS OF PATH LINKING BETWEEN OD PAIRS....................................94 TABLE 7.3 TRANSPORT COST BY MODE AND SHIPMENT SIZE.................................95 TABLE 7.4 COST OF DAMAGE DURING TRANSIT BY MODE AND COMMODITY...............................................................................................96 TABLE 7.5 COEFFICIENTS OF INDEPENDENT VARIABLES FOR MEAT/SEAFOOD...................................................................................................97 TABLE 7.6 COEFFICIENTS OF INDEPENDENT VARIABLES FOR NATURAL SANDS.................................................................................................97 TABLE 7.7 TOTAL UTILITY OF THE BEST ALTERNATIVES FOR MEAT/SEAFOOD...................................................................................................98 TABLE 7.8 TOTAL UTILITY OF ALTERNATIVES FOR NATURAL SANDS.................99 TABLE 7.9 COMMONALITY FACTORS FOR EACH PATH............................................100 TABLE 7.10 RESULTS OF THE ASSIGNMENT FOR N = 10............................................101 TABLE 7.11 RESULTS OF THE ASSIGNMENT FOR N = 100..........................................102 TABLE 7.12 RESULTS OF THE ASSIGNMENT FOR N = 1000........................................103 TABLE 7.13 ANALYSIS OF RESULTS OF THE ASSIGNMENT FOR N = 10, 100, 1000...................................................................................................106 x TABLE A1.1 DEVELOPMENT OF RANDOM UTILITY THEORY AND ORIGINS OF MNL MODELS................................................................134 TABLE A1.2 APPLICATION OF RANDOM UTILITY THEORY TO TRAVEL DEMAND ANALYSIS.......................................................................................135 TABLE A2.1 FIRST LEVEL (2DIGIT) STCC DEFINITION.............................................136 TABLE A2.2 FIRST LEVEL (2DIGIT) SCTG......................................................................137 1 CHAPTER 1 INTRODUCTION The demand for goods has increased steadily over the past half century and a costeffective freight transportation system has become an integral ingredient of a thriving national economy. Only recently, freight transportation has been systematically analyzed and planned and is comparatively new compared to passenger transportation planning. A critical part of freight transportation analysis and planning is freight transport modeling which is used to forecast and predict behaviors of actors (e.g. freight shippers and carriers) in freight transportation system as well as evaluate related policies and measures. However, the complexity of freight transportation models is far beyond that of passenger transportation. As indicated by Chiang et al. (1980), “In modeling freight transportation systems, models have been developed by researchers from many disciplines using many different approaches in an attempt to solve many different problems. This is just one indication that freight transportation involves complicated decisionmaking processes.” Other factors contributing to the complexity of freight transportation modeling include variables affecting freight movement patterns, for example, locations, range of transported commodities, characteristics and nature of raw materials and end products, manufacturing operations and demand variation and pricing (Ortuzar and Willumsen 1990).  2 In addition, decisionmaking processes in freight transport also relate to many interdependent actors such as suppliers, manufacturers, consumers, shippers (owners of goods who select modes of transport), carriers (owners of transport services who select transport routes) and government as infrastructure providers (Harker 1987 and Tavasszy 1996). 1.1 TAXONOMY OF FREIGHT TRANSPORTATION MODELS Freight transportation models can be generally classified into two main categories: operational models for short to mediumterm decision making and strategic models for longterm decision making. The details and differences between these two types of freight transport models are as follows (Kristiansen and Petersen 2002, Tavasszy et al. 2000). Operational Models – This type of freight transportation models is at the firm level and usually applied for optimization purposes. Compared to strategic models, operational models are closer to actual decision making and more detailed in their description of logistic activities. Besides, a multitude of data is potentially available for modeling at the operational level. Some examples of issues to be analyzed by this type of models are change in cost structure, change in transport market, weight and dimension of vehicle, locations of distribution centers and fleet and crew arrangement. Strategic Models – In contrast to operational models, a strategic model is an aggregation of firmrelated flows. Strategic models are usually descriptive in nature and used to obtain insight into the impact of freight flows on the infrastructure network for longterm planning purposes rather than to optimize decisionmaking processes as required by operational models. Strategic models can be used to analyze, for instance, effects of 3 transport policy measures on longterm patterns and modal distribution, effects of specific transport infrastructure investment projects on traffic pattern/modal distribution and socioeconomic and environmental impacts as well as assessment of transport network development plans. Some examples of policy questions to be analyzed by strategic models are (i) what is the influence of central distribution on transport pattern and mode share, (ii) what is the competition between ports and (iii) what consequences will multimodal transport policy have on utilization of the different transport networks. Most strategic modeling concepts applied in freight transportation have originally been developed based on the conventional fourstep sequential model approach widely applied in passenger transportation. However, the context of fourstep model in freight transportation is quite different from passenger transport as follows; (De Jong et al. 2004) Production and attraction: In this step, the quantities of goods to be produced in various origin zones and the demand for goods that are attracted/consumed in various destination zones are determined (the marginal of origindestination (OD) matrix). The output dimension is tons of goods. In intermediate stages of the production and attraction models, the dimension of freight flows could be converted to monetary units or vehicle units i.e. number of trucks used to transport goods. Distribution: The flows for each commodity type transported between each origin and destination (each cell of the OD matrix) pair are determined. Modal split: The allocation of the commodity flows to modes, for example, highway, rail and waterway. 4 Flow Assignment: Freight flows are assigned onto the network of each mode in this step. Flows could be assigned directly in tons onto the network or converted into vehicleunits before being assigned onto the network. 1.2 DISAGGREGATE DISCRETE CHOICE MODELS Aggregate demand or first generation transport models are either based on observed relations for groups of travelers/actors in the system or on average relations at a zonal level. Disaggregate demand or second generation models are based on observations of individual actors in the system, therefore, enable more realistic models. Spear (1977) summarized some advantages of disaggregate models as follows. Disaggregate models are based on theories of individual behaviors, therefore, an important advantage of the disaggregate models over conventional aggregate ones is that it is more likely that disaggregate models are stable and transferable in time and space. Since disaggregate models are based on individual data, they require fewer data points and are less likely to suffer from biases due to correlation between aggregate units. Disaggregate models are stochastic and they yield the probability of choosing each alternative rather than indicate which one is selected. Disaggregate models can have explicitly estimated coefficients and allow any numbers and specification of the explanatory variables while generalized cost function in aggregate models is generally limited to only costrelated variables (e.g. travel cost and travel time) and fixed parameters. This implies that disaggregate 5 models are a more flexible representation of policy variables. Besides, coefficients of explanatory variables directly reflect the relative importance of each attribute. The details of disaggregate discrete choice theory, derivation of models and some important properties are described in Appendix A1. 1.3 PROBLEM STATEMENT AND NEED FOR THE RESEARCH Freight transportation activities in the present circumstances are closely related to a larger context of logistics decisionmaking (e.g. inventory policy, warehouse and distribution center locations). Nowadays, commercial companies do not only consider transportation as the immediate cost of moving goods from one place to another, but they view transportation process as a part of the whole logistic concept. According to this concept, capital requirements related to easy and fast market access might be more important than the direct transport costs. Firms are also aware of the importance of response times and reliable delivery. Nevertheless, this logistic requirement is not relevant to all types of goods. For example, bulk products like crude oil and coal are commodities that can be kept in storage for relatively long periods. In this case, the direct transport costs are still of major importance for the choice of mode and transport route (Kristiansen and Petersen 2002). Most existing freight transport demand models used for mode and route selection still focus only on the direct factors e.g. transport cost and transit time. However, in practice, as described previously, other important logistics factors could also affect the transportation decisions of shippers and carriers. Therefore, this thesis effort identifies these factors and studies how they can be incorporated into freight transport demand models in order to improve the forecasting capability of the model. Conceptually, one 6 possible way to do so is by the application of discrete choice approach to mode and route selection models. Although, it is possible in principle, based on the review of freight transportation models both in the U.S. and in Europe and a study by De Jong et al. (2004), disaggregate behavioral models are not common in freight transportation models. Commodity characteristics are a key factor in decision making relating to transport mode and route selection. Further, with the application of the discrete choice model, commodities in the same group are assumed to have similar transportation behaviors. Therefore, a part of this research is also devoted to the classification of commodity groups. 1.4 OUTLINE OF THE THESIS The rest of this thesis document is organized as follows. Chapter 2 includes a review of freight transportation models in practice, modal split and flow assignment models, multimodal network models, classification of commodity groups based on logistical requirements, application of discrete choice models in transport mode and route selection, and revealed and stated preference survey for parameter estimation. This is followed by Chapter 3, the research statement, which explains the research goals, objectives, scope and limitations and contributions to the field of freight transportation demand modeling. Chapter 4 presents the research approach and enumerates the steps in order to accomplish the objectives. Chapter 5 is devoted to the review and reclassification of STCC and STCG commodity groups based on commodity characteristics and logistical requirements. The mathematical models based on the network clustering concept are used to classify commodities into logistical families. Chapter 6 presents potential factors, besides direct transport cost and transit time, for freight transport mode 7 and route selection model. The structure of discrete choice model based on the identified potential factors for freight transport mode and route selection is also proposed in the chapter. Finally, in Chapter 7, the deployment of the proposed discrete choice model for multicommodities multimodal stochastic network flow assignment is explored. First, the conventional stochastic user equilibrium assignment model is studied. Then, the application of the proposed discrete choice model to the conventional assignment model is presented. A heuristic algorithm to assign freight flows onto multimodal stochastic and congested network based on the proposed discrete choice model is developed. A proofof concept implementation of the heuristic algorithm for a simplified network with multiple origindestination pairs and commodities is also presented. 8 CHAPTER 2 LITERATURE REVIEW This chapter begins with the review of freight transportation models developed both in the U.S. and Europe. Then, algorithms used for transportation modal split and flow assignment are reviewed. In section 2.3, multimodal network models and the methods to develop such models are described. The concept of classifying commodities according to their logistical requirements is presented in section 2.4. The application of disaggregate discrete choice models in the transportation context, especially in mode and route selection is summarized in section 2.5. The chapter ends with a review of revealed and stated preference survey for parameter estimation in discrete choice models. 2.1 REVIEW OF FREIGHT TRANSPORTATION MODELS IN PRACTICE In this section, freight transportation models developed both in the U.S. and Europe are reviewed. 2.1.1 Freight Transport Models in Europe 2.1.1.1 Strategic Model for Integrated Logistic Evaluation (SMILE) – SMILE is a strategic model for freight transport and logistics used in Netherlands. 9 Although it is a strategic model describing aggregate level of national freight transport, it has the ability to incorporate the relationship among production, inventory and transportation at disaggregate level. This relationship can be represented and analyzed in the model by a threelevel modeling approach as follows; (Tavasszy et al. 1998) Module of production, sales and sourcing – the location pattern of both production and consumption are established by using Make/Use tables. After the volume and nature of production and consumption at different locations are determined, the spatial distribution from production sites to consumption sites resulting from comparative price differences and the resistance of geographical, organizational and institutional differences are determined. Inventory Module – The main function of this secondlevel model is to link trade relations to transport relations by considering warehousing services. New groups of commodities are formed based on the product and market characteristics such as the value of products per cubic meter, packing density and perishability and so forth. Optimal locations of distribution centers are determined by using criteria such as lead time, closeness to activity centers and available modes of transport. Then conditional on the locations of distribution centers, multinomial logit models are used to assign flows to alternative channel types. Transportation Module – Freight flows in each distribution channel resulting from the Inventory Module are assigned onto the multimodal network of six modes using aggregate approach to find the optimal route (AllorNothing) with cost minimization. 2.1.1.2 Network Model for Norwegian Freight Transportation (NEMO) – The conventional fourstep model including generation, distribution, modal split and 10 assignment is applied directly in NEMO. Only eleven groups of products are included in the model. The network in the model is a multimodal network, therefore, allowing for simultaneous mode and route analysis. The transport modes included in the network are road, rail, seaborne transport, air freight and pipeline. The assignment algorithm in the model is aggregate allornothing based on cost minimization. The cost structure is comprised of reload costs when mode transfer occurs, qualitative costs (e.g. risk of delay and travel time) and truck, ship and rail operation costs (Hovi and Vold 2003). 2.1.1.3 The Walloon Freight Transport Model for Belgium (WFTM) – WFTM also applies the conventional fourstep model. The base year OD matrices in 1995 of ten commodity classes were constructed as the part of freight flow generation and distribution steps. In this model, modal split and flow assignment are also simultaneously analyzed on its multimodal network using the system optimal algorithm (e.g. an assignment algorithm aiming to minimize the total generalized cost in the network). An interesting concept of WFTM is the utilization of a virtual network to represent multimodal network in physical configurations and operational characteristics. The operational characteristics modeled in WFTM are loading, transshipment, waiting and handling (Jourquin and Beuthe 2000 and Jourquin and Limbourg 2006). The details of the virtual network will be described in more detail again in section 2.3 of the literature review. 2.1.1.4 National Freight Model System for Sweden (SAMGODS)  Base year OD matrices in the year 1997 were constructed for six main groups, split into bulk/general, cargo, high/low density and high/low value. The basis for the model is a set of inputoutput tables. The model calculates forecasts of production, import, export, inputs to production and consumption in monetary units for 31 different sectors. For the network 11 model in SAMGODS, a multimodal network of road and rail modes is used and the flow assignment is analyzed by the system optimal approach which minimizes the generalized cost of the whole network (MEP and SWP 2002 and De Jong et al. 2002). 2.1.1.5 Decision Support System for Transport Policies of Italy – Here also the concept of the fourstep modeling approach is used. Generation and distribution models are represented by multiregional input/output models. In this Italian freight transport model, modal split is separated from the assignment step. Disaggregate discrete choice model is used to forecast mode share among seven modes: road by own truck, truck of carrier for a single shipment or truck of carrier under contract, traditional rail, combined rail (contained, swap bodies or semitrailers), shipper by road, shipper by rail, shipper with mode chosen by himself. The segmentation of commodities believed to influence mode choice such as perishable, consumer and capital goods are classified. The assignment stage takes the multimodal OD matrices and assigns them to networks. For road, random utility path choice models are used (MEP and SWP 2002). 2.1.2 Freight Transport Models in the U.S. For the statewide freight transportation models in the U.S., the structures of the models are quite similar. The unit of flow is typically number of trucks for vehicular flows and tons for commodity flows. The main sources of OD data of freight are Commodity Flow Survey (CFS) (Bureau of the Census and Bureau of Transportation Statistics, 1997) and TRANSEARCH database (http://www.globalinsight.com/Transearch). Most of the models use the fourstep sequential modelling approach. Details of the models for some states are briefly described below; 12 2.1.2.1 Statewide Freight Transport Model for Indiana – There are 145 Traffic Analysis Zones (TAZs) in the model representing 92 counties in Indiana and 53 more for other 47 states (not including Hawaii and Alaska) in the U.S. CFS (Bureau of the Census and Bureau of Transportation Statistics, 1997) is used as the main source for the OD flow matrix. Two modes of transport, namely, truck and rail, are considered in the model. The modal split model uses the CFS modesplit ratio in 1993 for the future years. The commodity flows in tons for both rail and truck are converted to vehicle trips and assigned onto their own network by AllorNothing approach (Horowitz and Farmer 1999). 2.1.2.2 Statewide Freight Transport Model for Wisconsin – There are 132 Traffic Analysis Zones (TAZs) in the model representing 72 counties in Wisconsin and 60 more for other 47 states in the TRANSEARCH database is used as the main source for OD flow matrix. Highway, rail, waterway and air transport modes are considered in this model. Aggregate logit model for all commodities is used to split mode share. Based on the information from vehicle inventory and usage survey, payload factors are used to convert commodity tons to truck trips and conversion factors are used to convert annual truck traffic to daily volumes. Daily truck volumes are assigned onto the network by using multiclass User Equilibrium assignment with preloaded passenger volumes. However, commodity flows in tons are assigned onto the rail network by using Allor Nothing algorithm (Center for Urban Transportation Studies, 1999 and Transportation Systems Design Graduate Students, 2006). 2.1.2.3 Statewide Freight Transport Model for Iowa – There are 145 TAZs in the Iowa statewide freight model representing 99 counties in Iowa and Business Economic Areas 13 (BEAs) in other states. Only three main commodity groups (e.g. grain, meat products and machinery) are considered important and included in the model. The main source of OD freight flows is TRANSEARCH. Highway transport by trucks is only the transport mode considered in the model. Commodity flows in tons are assigned onto the network using AllorNothing algorithm (Souleyrette et al. 1996). 2.1.2.4 Statewide Freight Transport Model for Oklahoma – There are 204 TAZs in the model representing 77 counties in Oklahoma and 127 business areas in other states. The main source of OD freight flows is the Freight Analysis Framework (FAF2.2) developed by Federal Highway Administration, US Department of Transport (Federal Highway Administration, 2006). The total commodities are classified into 43 groups based on Standard Classification of Transported Goods (SCTG). The ratios among transportation modes; highway, railway and waterway in 2002 are used for mode choice in the future years. The commodity flows transported by railway and waterway are assigned on their own network using shortest travel distance algorithm. For highway mode, the flows are assigned onto the network by using shortest travel distance, shortest travel time, user equilibrium and system optimal algorithms. Furthermore, a mathematical programming model is underdevelopment in order to assign highway freight flows to minimize the total system travel time by simplifying the travel timeflow relation to be piecewise linear. All capacitated assignments are analyzed after preloading passenger flows onto the network (Ingalls et al. 2007 and Ingalls et al. 2009) A summary of the reviews of the above freight models, especially with respect to the aspects of modal split, flow assignment, multimodal consideration and inclusion of logistical requirements is shown in Table 2.1. 14 Table 2.1: Summary of the Review of Freight Models in Europe and the U.S. Model Name Modal Split Flow Assignment Multimodal Consideration Logistical Requirements Consideration Strategic Model for Integrated Logistics Evaluation (SMILE) Not included Not included Multimodal network assignment by optimal route (AllorNothing) algorithm based on cost minimization Included in the model based on commodity classication and inventory module considering optimal distribution channel Norwegian National Freight Model System (Nemo) Not included Not included Multimodal network assignment by optimal route (AllorNothing) algorithm based on cost minimization Not included The Walloon Freight Transport Model for Belgium (WFTM) Not included Not included Multimodal network assignment by System Optimal to minimize costs of the total network Included in the model based on virtual network considering logistical activities at transfer/transshipment facilities Swedish National Freigt Models System (SAMGODS) Not included Not included Multimodal network assignment by System Optimal to minimize costs of the total network Included in the model based on commodity classication as bulk/general, cargo, high/low density and high/low value Decision Support System for Transport Policies of Italy Disaggregate discrete choice model Random utility path choice model for road transport Not included Included by segmentation of commodities believed to influence mode choice: perishable, consumer and capital goods Statewide Freight Tranport Model for Indiana Ratio of mode share from CFS database Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Tranport Model for Wisconsin Aggregate logit model Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Tranport Model for Iowa Only truck mode is considered Assignment by optimal route (Allor Nothing) algorithm based on cost minimization Not included Not included Statewide Freight Transport Model for Oklahoma Ratio of mode share from FAF database for highway, railway and waterway modes Railway and waterway flows by shortest distance and capacitated assignment for highway Not included Not included 1. Freight Transport Models in Europe 2. Statewide Freight Transport Models in U.S. 15 2.2 TRANSPORTATION MODAL SPLIT AND FLOW ASSIGNMENT As described in Chapter 1, modal split and flow assignment models are the last two steps within the conventional fourstep model. The main functions of modal split and flow assignment models, in the context of freight transportation, are to allocate commodity flows to different modes and assign flow in each mode onto the corresponding transportation network. 2.2.1 Modal Split models – Many algorithms have been used to represent decisionmaking for mode selection ranging from an easy and coarse method to a more complicated and detailed one. The easiest method is to use empirical data to develop the ratios of sharing among modes. Another possible way is to construct a diversion curve describing the relationship between proportions of flows by mode against the difference of some factors (e.g. cost and time difference) between any two modes (Ortuzar and Willumsen 1990). De Jong et al. (2004) also provided a taxonomy of modal split models as follows. Elasticitybased models reflect the effect of changing a single variable (e.g. cost of some modes) to the proportion of mode selection. This type of model is usually used for strategic evaluation or for a quick first approximation. Elasticities can be derived from other models such as aggregate mode choice model or from expert knowledge. Aggregate modal choice model – The models are generated from the concept of entropymaximization. In entropymaximization, a system is made up of a large number of distinct elements and can be classified into three main levels: micro, meso and macro. The concept postulates that there are numerous and different states in micro and meso levels which produce the same state in macro level and assumes that 16 all micro states consistent with our information about macro states are equally likely to occur. The derived aggregate models are similar to disaggregate discrete choice models in logit family. However, explanatory variables are at the aggregate level and parameters are estimated based on zonal/interzonal information. The main disadvantage of aggregate modal choice model is the lack of insight into causality of mode choice from the individual perspectives. Neoclassical models – the models are derived from the economic theory of the firm. For a cost function, transport services are considered as one of the inputs. The explanatory variable in the models is the budget share of some modes in the total cost. This type of model is difficult to combine with the conventional fourstep model because the share in the transport volume is the relevant variable. Direct demand models – These models can generate the number of trips by modes directly (unlike generated in market share forms by other types of models). Because the models directly generate number of trips by each mode, it is quite separate from the product/attraction and distribution steps in the fourstep model and make it difficult to incorporate it into the fourstep framework. 17 Disaggregate modal choice models – This type of models is derived from the concept of random utility theory. The model structures of logit family are similar to those of aggregate modal choice models. To estimate models’ parameters, behavioral data from individual units, e.g. shipper and carrier in the freight transport context, are required. More details of disaggregate modal choice models are presented in Appendix A1 and section 2.5 in this chapter. A summary of freight transport modal split models is presented in Table 2.2 (De Jong et al. 2004). 2.2.2 Flow Assignment Models – Flow assignment models can be classified with respect to supply characteristics and demand assumptions underlying the models. For supply characteristics, a network can be considered congested when link performances (e.g. travel time and cost) depend on link flows or uncongested, if link performances are Type of Model Advantages Disadvantages Modal share ratios/Diversion curve Quickest and easiest to develop based on historical data Lack of supporting theory and concerns about predicting capability Elasticitybased models Very limited data requirements and quick to develop Elasticities may not be transferable; measure only impact of single variable; no synergies Aggregate modal split models Limited data requirements Little insight into causality based on real behaviors of individuals and limited scope for policy effects Neoclassical models Limited data requirements and supported by theoretical basis Hard to integrate into fourstep models Disaggregate modal split models Supported by theoretical basis, able to include casual variables and policy sensitive measures based on real behaviors of individuals Need disaggregate data from individual (shippers/carriers) for parameter estimation Table 2.2: Summary of Freight Transport Modal Split Models Source: National and International Freight Transport Models: An Overview and Ideas for Future Development by De Jong et al. (2004) 18 fixed and independent of link flows. For a congested network, costflow or travel time flow curves are used to represent the relationship between link flows and link performance. Costflow relationships proposed by researchers are as follows; Smock (1962) for Detroit Study; exp (V/C) 0 t = t Overgaard (1967) (V/C) 0 t = t Bureau of Public Roads (1964) in the U.S. ] [1 (V/C) 0 t = t + where V is link volume in passenger car equivalent unit/hour. C is link capacity in passenger car equivalent unit/hour t is congested travel time of the link 0 t is freeflow travel time of the link and b are parameters On the demand side, if it is assumed in the model that all users perceive network performances in the same way as unique values and have no personal preference, the model can be considered deterministic. If the model allows the possibility of perceiving network performances differently and users can have personal preference, the model is stochastic. In the case of deterministic assignment, all users of the link are assumed to perceive the link cost in same way as the mean link cost. However, in the case of stochastic assignment, users of a link are allowed to differently perceive the link cost based on the assumption of link cost distribution. 19 In addition, flow assignment models can be aggregate if they are derived from zonal/interzonal data or disaggregate if they are derived based on individual data. Some algorithms developed for flow assignment models are as follows (Ortuzar and Willumsen 1990); Allor–Nothing  This is the simplest route choice and assignment method. The algorithm is for deterministic uncongested network flow assignment. The algorithm is probably reasonable for sparse and uncongested networks e.g. intercity/interstate networks. However, for networks in urban areas where there are high congestion effects, this model may not be realistic. User Equilibrium – This model can be considered deterministic congested network assignment. The model is based on the Wardrop’s first principle (1952) postulated that “Under equilibrium conditions traffic arranges itself in congested networks in such a way that no individual trip maker can reduce his/her path costs by switching routes and all used routes have equal and minimum costs while all unused routes have greater or equal costs”. It means that this algorithm minimizes generalized costs of each individual user in the system. System Optimal – The model is based on the Wardrop’s second principle (1952) postulated that “Under social equilibrium conditions, traffic should be arranged in congested network in such a way that the generalized costs of the whole system are minimized”. In contrast to the user equilibrium reflecting behaviors of individuals trying to minimize their own costs, this algorithm is oriented towards transport planners who try to minimize total system costs. 20 Simulationbased Methods – This approach was proposed by Burrell (1968). It can be considered as an approach for stochastic uncongested network assignment. In this approach, for each link in a network, link cost is separated into objective or engineering costs as measured by a modeler or the observer of the system and subjective cost as perceived by real users in the system. Therefore, it can be assumed that users perceive link cost as random variable with the engineering cost as the mean and distributed according to some known distributional form e.g. Uniform or Normal. To assign flows onto the network, total flows are divided into N segments, Monte Carlo simulation used to generate random generalized cost for each segment, then, flows in each segment are assigned onto the network based on AllorNothing approach. Singlepath Proportional Stochastic Methods – This is another stochastic uncongested network assignment algorithm proposed by Dial (1971). The method is based on a loading algorithm which splits trips arriving at a node between all possible exit nodes. For example, in Figure 2.1, flows from origin I to node B will be split into the flows from Ai to B by splitting factors which can be derived from the extra cost incurred in traveling from the origin to node B via node Ai rather than only via the minimum cost route. 21 Discrete Path Choice Models – There are some applications of discrete choice models in the context of transportation route choice analysis. The applications usually are for stochastic uncongested network assignment. This type of models is discussed in more detail in section 2.5. Stochastic User Equilibrium – In this case, the transportation network is considered for both congested and stochastic aspects. Performances of a link depend on link’s flows and can be perceived differently by each user on the links. The assignment problem can be formulated as a mathematical program and solved to optimality using an algorithm called the Method of Successive Averages (MSA) (Sheffi, 1985). The details and algorithm for stochastic user equilibrium are discussed in Chapter 7. A summary of the algorithms used in flow assignment according to supply and demand characteristics is shown in Table 2.3. I J A4 A2 A1 A3 A5 B Figure 2.1 Links Feeding Trips into a Node Source: Modelling Transport by Ortuzar and Willumsen (1990) 22 Table 2.3: Summary of Algorithms used for Flow Assignment Uncongested Network Congested Network Deterministic AllorNothing User Equilibrium, Systetm Optimal Stochastic Simulationbased, Singlepath proportional, Stochasitic Methods and Discrete Choice Model Sotchastic User Equilibrium 2.3 MULTIMODAL NETWORK MODELS Recently, multimodal transportation has become increasingly important since it can provide greater efficiency and cost savings than monomodal transport. According to the 1997 Commodity Flow Survey (CFS) (Bureau of the Census and Bureau of Transportation Statistics, 1997), tens of thousands of intermodal shipments were operated out of the total shipments of around five million tons (Bureau of the Census and Bureau of Transportation Statistics 1997). Southworth and Peterson (2000) proposed a method to construct a multimodal network from monomodal networks. In their work, three main transportation networks; truck, rail and waterways, are combined together into an intermodal network. Transfer terminals among modes (e.g. truckrail, truckwaterways and waterwaysrail) are used as the connecting nodes in the intermodal network. The detailed National Intermodal Terminals Database (Middendorf 1998) is used for this purpose. Harker (1987), Crainic et al. (1990), Jourquin and Beuthe (2000) and Jourquin and Limbourg (2006) proposed the concept of virtual links to represent specific costs for particular uses of transportation infrastructure because a simple network does not provide an adequate basis for detailed analyses of transport and logistics operations. For example, trucks of different sizes and operating costs can use the same highway or at a terminal, a truck’s load can be transshipped on a train, bundled with some 23 others on a boat or simply unloaded as it reaches its station. Although these operations in the example use the same infrastructure, but the costs of each operation are different. Therefore, virtual links are needed for representing different operations/costs on the same infrastructure. An example of virtual network construction can be illustrated by the following example. Figure 2.2 shows a simple network of waterways consisting of 4 nodes and 3 links. Each link represents waterways that can support ships of 300, 1,350 and 600 tons respectively. Figure 2.3 shows the corresponding virtual network in which alternative operations are enumerated, for example, the shipment can be directly transported by 300ton ships from node a to node d or be transshipped at node b and c then transported to node d. This concept can be very useful for reflecting and integrating detailed logistical operations in freight transportation models. a1 b4 b3 c3 b2 c2 d2 c1 c4 c6 300t b1 b5 c5 d1 1350t 1000t 600t 600t 300t 300t Figure 2.3 Corresponding Virtual Network Source: Multimodal Freight Networks Analysis with NODUS: A Survey of Several Applications by Jourquin and Beuthe (2000). Figure 2.2 Simple Network of Waterways Source: Multimodal Freight Networks Analysis with NODUS: A Survey of Several Applications by Jourquin and Beuthe (2000). a b d 300t 1350t 600t c 24 2.4 CLASSIFICATION OF COMMODITIES BASED ON LOGISTICAL REQUIREMENTS To integrate logistical processes into freight transportation modeling, commodity classification is considered an important issue (REDEFINE 1999 and Tavasszy et al. 1998). This is especially true for mode and route selection steps. Basically, the classification should be based on handling characteristics and the fact that different commodity groups have different values of time. In project REDEFINE and SMILE, commodity groups are classified based on product characteristics and logistical requirements as follows; Bulk or general cargo Density (Kg/m3) Packaging density (packs/unit/ m3) Value (Euro/kg.) Use of distribution centers (Yes/No) Consignment size (small/medium/large) Value of time (low, medium, high) Demand frequency In the U.S., there are two main systems for classification of transported commodities: Standard Transportation Commodity Code (STCC) and Standard Classification of Transported Goods (SCTG). The details for each system are as follows (Federal Highway Administration 2006); Standard Transportation Commodity Code (STCC) – STCC system was developed in 1960s by a special committee of the Association of American Railroads (AAR). The 25 main purpose of the development was to serve users of AAR, particularly, North American Freight Railroads. The annual Railroad Waybill data, 1993 Commodity Flow Survey (CFS), and the first generation of the FAF all used the STCC coding system. There are 4 levels in the hierarchical system from 2digit to 5digit codes. Generally, the first four digits of the STCC represent the industry that produces the commodity, based on the Standard Industrial Classification (SIC) system. The fifth digit of STCC provides product classes within the producing industries. The last two digits of the STCC add commodity detail of particular interest to the railroads. A summary of the various 5digit STCC levels is presented in Table 2.4. The top level (2digit) STCC codes are listed in Appendix A2. Table 2.4: Levels of Classification in STCC Source: Report 4 (R4): FAF Commodity Classification by Federal Highway Administration (2006) Standard Classification of Transported Goods (SCTG)  The U.S. Department of Transportation, U. S. Bureau of the Census, Statistics Canada, and Transport Canada developed the SCTG to replace the STCC for the 1997 and subsequent CFS. The structure of SCTG is similar to that of STCC. There are also 5 levels of commodity codes in SCTG. At the most aggregated level (i.e., 2digit), the SCTG was designed to provide analytically useful commodity groupings for users that are interested in an overview of transported goods. With a small number of exceptions, categories in the 3digit level were designed to include goods for which significant product movements are expected to be recorded in both the United States and Canada. The 4digit SCTG Level Number of Categories Grouping (example) 2digit 37 Major industry classes (01Farm products) 3digit 182 Minor industry classes (012Fresh fruits or tree nuts) 4digit 444 Specific industries (0121Citrus fruits 5digit 1,202 Product classes (01214Oranges) 26 Level of Hierarchy Number of Categories Grouping (example) 2digit 42 Analytical overview 3digit 133 U.S.Canadian product groups 4digit 283 Transportation characteristics 5digit 504 CFS 2002 collection level categories were created to reflect industry patterns and transportation characteristics. The most detailed SCTG category, which is at 5digit level, is the collection level for the CFS. At this level, each category was designed to capture significant details that reflect industry patterns and transportation characteristics. Because most 4 and 5 digit SCTG categories primarily contain the products of only one industry, they can be associated with the SIC, as well as with the NAICS. This feature allows comparisons to be conducted with industry data, as well as other SICbased classifications such as the STCC system. The number of categories in each level of SCTG, as used in the 2002 CFS, is summarized in Table 2.5. The first level categories of the SCTG are listed in Appendix A2. 2.5 APPLICATIONS OF DISAGGREGATE DISCRETE CHOICE MODELS IN TRANSPORTATION The early applications of disaggregate discrete choice models in the area of transportation were made mainly for the binary choice (e.g. only two alternatives in a choice set) of travel mode. Some of these studies focused on the tradeoff between travel time and travel cost implied by travel demand models. Thereafter, further development of disaggregate discrete choice models in transportation was directed toward the choice of transport modes with more than two alternatives (multinomial discrete choice models) Source: Report 4 (R4): FAF Commodity Classification by Federal Highway Administration (2006) Table 2.5: Levels of Classifications in SCTG 27 and other transportrelated choice situations such as trip destination, trip frequency, car ownership, residential location and housing and routes selection (CRA 1972, BenAkiva 1973 and 1974, Brand and Manheim 1973, Richards and BenAkiva 1975 and Lerman and BenAkiva 1975.) For mode selection, Cascetta (2001) defined disaggregate mode choice models as the models used to simulate the probabilities of a user to select transport modes from an origin zone to a destination zone. Identification of modal alternatives in choice set depends on each transportation system under study. By using different heuristic approaches, trivially nonavailable alternatives can be eliminated such as choice of walking or bicycling can be excluded from the choice set of interurban transport system. For conditionally nonavailable alternatives, probability of including such alternatives in the choice set can be calculated, then probability of selecting alternatives in choice set can be identified as the conditional probability. Explanatory variables included in utility functions of mode choice models usually are levelofservice/performance attributes of each alternative and individual’s socioeconomic attributes. Examples of performance attributes in mode selection models could be travel time, travel cost, regularity of services and number of transfers. Socioeconomic attributes are characteristics of the decisionmaker in the system. This type of attributes is generic and not dependent on alternatives. Examples of socioeconomic attributes for mode choice models could be gender, age, family income and car ownership. Disaggregate mode choice models are quite common in transportation planning. Examples of passenger transport mode choice models are Warner 1962, Lisco 1967, Quarmby 1967, Lave 1969, Watson 1974, Rassam et al. 1971, 28 Ben Akiva 1973 and McFadden 1974. For freight transport mode choice model, some studies related disaggregate mode choice model are as follows (De Jong et al. 2004); Winston (1981): probit model for the choice between road and rail transport by commodity group in the U.S. Jiang et al. (1999): nested logit model on the French 1988 shippers survey. Nuzzolo and Russo (1995): mode choice model for the Italian national model. Fosgerau (1996): mode choice model on revealed and stated preference data. Reynaud and Jiang (2000): European freight model focusing on operating systems for rail developed with mode choice model on revealed and stated preference. Chiang et al. (1981): disaggregate model for selection of supplier, shipment size and transport mode. McFadden et al. (1985): disaggregate model for shipment and mode selection. Jovicic (1996): SP and RP interviews of Danish freight shippers and carriers to examine the importance of a number of different parameters describing their transport decision making. For path selection, disaggregate path choice models provide the probability of path selection of a user from an origin zone to a destination zone for a particular transport mode (Cascetta 2001). Disaggregate path choice models can be typified into two main categories: pretrip choice where the whole traveling path is chosen before starting the trip and pretrip/en route mixed choice where the route is chosen both before the starting and during the trip, for example, the selection of route for urban transit system with high frequency and low regularity. As mentioned in Appendix A1, the crucial parts of 29 disaggregate path choice models are identification and inclusion of path alternatives into the choice set. There are two ways to handle this issue as follows (Cascetta 2001): Exhaustive approach – The approach includes all elementary paths in the network into the choice set. It may generate a very large number of routes sharing many links. Selective approach – In contrast to the exhaustive approach, this approach identifies only some elementary paths in the choice set on the basis of heuristic rules. The heuristic rules used in path choice models depend on the application context. Some examples of heuristic rules for path choice alternatives are shown in Table 2.6. Another important issue for disaggregate path choice model development is the correlation among alternatives. It can be seen that alternatives in route choice models could be highly correlated because there are many alternative routes sharing links. In the case that the random part of the utility function is assumed to be Gumbel distribution and the discrete choice model is called "logit choice model", this correlation issue could introduce complexity (the derivation of logit choice model and its property are presented Table 2.6: Criteria for Path Feasibility Selection Criteria Specification Topological A path is feasible (Dial efficient) if each link "goes away" from the origin and/or "move towards" the destination. Comparison of costs Paths with a generalized cost not exceeding by more than X times of minimum cost. Progressive The first n minimum generalized cost paths. Multiattribute Minimum paths with respect to various attributes (usually the relevant perfomance variables such as travel time, monetary cost, motorway distance and so forth). Behavioral Paths excluding behaviorally unrealistic link sequences (e.g. repeated entrances and exits for the same motorway). Distinctive Paths overlapping for no more than a given percentage of their length. Source: Transportation Systems Engineering: Theory and Methods by Cascetta (2001) 30 in Appendix A1). The reason is that there is a key assumption to derive the logit choice models that the disturbance term in the utility function is identically and independently distributed meaning that all alternatives must be independent. The explanatory variables generally applied in path choice models are path performance or levelofservice such as travel time, travel cost, number of transfers, service frequency, level of congestion and so forth. Disaggregate path choice models are not widely applied in passenger transport and, according to the literature review, they are not common in freight transport either. Limited application of disaggregate path choice model may stem from their complexity and the difficulty in data collection compared to the aggregate model. Examples of path choice models in passenger transport are as follows; BenAkiva et al. (1984): Discrete path choice model for passenger transport by defining choice sets of labeled paths to transform a large number of physical routes into a smaller number of labeled routes. Dial (1971): A Probabilistic Multipath Traffic Assignment Model approach to include reasonable paths into the choice set. Paths included in the choice set should be composed of links that would not move the traveler farther away from his/her destination. Burrell (1968): Stochastic models seeking to account for variations in drivers’ perceptions of travel times or costs by means of a probability distribution for perceived link performances. Hidano (1983): Proposing individuals’ route plan in a hierarchical fashion starting at the lowest level in the road network, proceeding up a hierarchy and then going down the hierarchy in the vicinity of their destination. 31 2.6 REVEALED AND STATED PREFERENCE SURVEY FOR PARAMETER ESTIMATION Disaggregate discrete choice models are data dependent meaning that the quality and accuracy of the models largely depend on the data used for the models’ parameters estimation. Before the mid of 1980s, parameter estimation of discrete choice models was dominated by revealed preference data. Revealed preference (RP) data is based on real behavior observed in an actual system. However, revealed preference data has some limitations as follows (Ortuzar and Willumsen 1990); Observations of actual choices may not provide sufficient variability for constructing good models for evaluation and forecasting. For example, travel time and travel cost may be highly correlated and may be very difficult to separate their effects in model estimation. The observed behavior may be dominated by a few factors making it very difficult to detect the relative importance of other variables, especially for qualitative variables. For the alternatives which are entirely new such as completely new mode of transport, it is very difficult to collect data by using revealed preference technique. According to these reasons, a new technique was required to fill the gaps. By the end of 1970s, stated preference (SP) technique that originated in the field of market research was proposed to overcome the limitations of revealed preference data. The SP technique offers a way of experimenting with choices directly, thus solving some of the above limitations. SP technique estimates the parameters of discrete choice model based on an analysis of the response to hypothetical choices which can cover a wider range of attributes and conditions than the real behavior as applied in RP technique. By using SP 32 survey, individuals are asked about which choice they would make in one or more hypothetical situations. Cascetta (2001) summarized the advantages of SP over RP technique as follows; SP technique allows the introduction of choice alternatives not available in the present. They can control the variation of relevant attributes outside the present range to obtain better estimates of the relative coefficient. They can introduce new attributes and their coefficients, especially, qualitative ones such as convenience and safety. However, these advantages come with the tradeoff of introducing some distortion in the results because individuals respond based only on the hypothetical and not the real situation. Besides, it is possible that the stated choices are presented in unrealistic ways for example, some attributes presented to individuals might be missing or there may be fatigue and justification bias effects and these situations can further increase the distortion in the models. Hensher (1994) summarized the steps in implementing SP survey as follows; Task 1: Identification of the set of attributes – There may be a large number of attributes to be included in an alternative, therefore, it is required to decide early on which attributes should be included. Task 2: Selection of the measurement unit for each attribute – In most cases the units of attributes are unambiguous, for example, hours for travel time and dollars for travel cost. However, for some qualitative attributes such as convenience, reliability and safety, the units of attributes should be clearly identified to avoid possible confusion of respondents. 33 Task 3: Specification of the number and magnitudes of attributes – The levels of attributes are specified in this step to measure variation of each attribute in SP survey. For example, the levels of travel cost can be divided into three levels: high, medium and low cost. The value of attributes in each level can be identified based on the existing value in an actual system. Task 4: Statistical Design– In this step, attributes and their levels are combined to construct hypothetical situations. The design of these hypothetical situations can be based upon the experimental design principle widely used in statistics. Full factorial or fractional factorial design can be applied. Task 5: Translation of statistical design into a set of questions for data collection– The experimental design in task 4 is translated into a set of question to ask respondents. One should be careful about setting up the questions used to asked respondents. McFadden and Leonard (1992) conducted the tests of SP methods and compared results from alternative SP experiments different in response format, question phrasing and information provided to the respondent. They found great sensitivity of the results according to these factors. As mentioned earlier, both RP and SP techniques have their own advantages and disadvantages. RP data provides the real behaviors of individuals in an actual situation while SP data allows the introduction of nonexisting alternatives and control of attributes’ variation. Therefore, to improve the models’ estimation, one may be required to combine both RP and SP data together. This process was originally proposed by Morikawa (1989). The data enrichment paradigm (Louviere et al. 2000) according to Morikawa’s concept can be depicted in Figure 2.4. For this paradigm, RP data are viewed 34 as the standard of comparison among alternatives and SP data are seen as useful only to the extent that they lessen undesirable characteristic of RP data by mainly using tradeoffs among variables from hypothetical experiments. Respondent SP Data SP Equilibrium SP Tradeoffs RP Data RP Equilibrium RP Tradeoffs Respondent Figure 2.4 Data Enrichment Paradigm Source: Stated Choice Methods: Analysis and Applications by Louviere et al. (2000) 35 CHAPTER 3 STATEMENT OF RESEARCH The overall goals of this thesis effort were (i) to improve freight transportation modeling by making it more realistic as well as enabling it to represent firmlevel logistical operations aspects and (ii) to investigate how to properly apply disaggregate discrete choice models in the context of freight transport mode and route selection. 3.1 RESEARCH OBJECTIVES According to the research goals mentioned above, the objectives of this thesis can be enumerated as follows; Objective 1: To perform a review of literature related to freight transportation modeling, especially focusing on modal split, flow assignment, multimodal network consideration and inclusion of logistical requirements. Objective 2: To identify criteria for commodities’ logistical characteristics (e.g. time value, physical configuration) and requirements (e.g. bulk/containerized, consignment size) and systematically classify commodity groups according to those criteria. Objective 3: To identify potential explanatory variables which should be included in the utility function when shippers/carriers decide on freight transport mode and routes selection and specify a proper structure of disaggregate discrete choice models (e.g. linear in parameter) based on the identified variables. 36 Objective 4: To investigate the changes to the conventional mathematical program for network flow assignment for applying the proposed disaggregate discrete choice model with the identified explanatory variables. Objective 6: To test the proposed discrete choice model and the network assignment algorithm on a simplified multimodal network with multiple OD pairs and commodities. 3.2 RESEARCH SCOPE AND LIMITATION Because of time and resource constraints, the scope of this research is limited as follows; This research considers only mode and route selection models and excludes flow generation and distribution steps by assuming that freight flows between each origindestination pair are given in the model. The research considers only three modes of transport: (highway, railway and waterway). Each mode can have its own capacity, however, only travel time on the highway network is considered to be affected by congestion. Many commodity groups with different utility functions and different origindestination pairs can be assigned onto the network simultaneously. The multimodal network used to test the developed algorithms is simplified from the real network with link performance data (e.g. travel cost, travel time, travel time variation and so forth). Because parameter estimation is outside the research scope, the parameters of discrete choice model are assumed in this experiment. The mathematical model for network flow assignment considered in this research is stochastic user equilibrium model. 37 3.3 RESEARCH CONTRIBUTION The purpose of this thesis is to contribute to the improvement of strategic freight transportation modeling, especially in mode and route selection tasks. The main contributions can be described as follows; A comprehensive review of literature related to freight transportation modeling, especially, in the aspect of multimodal network and modeling based on logistical requirements and capacity/congestion effect. The improvement of the criteria used to classify commodities by considering logistical characteristics and requirements and the application of clustering approach for the commodity classification. Identification of discrete choice model structure and explanatory variables to be included in the model as well as the algorithm for applying the proposed discrete choice model for stochastic congested network flow assignment. 38 CHAPTER 4 RESEARCH APPROACH To complete the objectives described in Chapter 3, this chapter briefly explains the tasks that must be accomplished. Four main tasks were carried out as part of this research; Identification of criteria for classification of commodity groups based on commodity characteristics and logistical requirements and creation of the commodity groups based on the criteria developed and mathematical models based on clustering approach. Identification of disaggregate discrete choice model structure and explanatory variables to be included in the model. Investigation of the impacts and changes due to the proposed discrete choice model to the conventional mathematical program for stochastic user equilibrium assignment and development of an algorithm for applying the proposed discrete choice model to stochastic congested network flow assignment Proofofconcept implementation of the assignment algorithm on a simplified multimodal network with the assumed models’ parameters. The flow chart of all tasks in the thesis are depicted in Figure 4.1 39 4.1 EXPLANATION OF THE RESEARCH TASKS 4.1.1 Identification of Logistical Requirements and Classification of Commodities Groups Based on the fact that each commodity has its own logistical characteristics and requirements, decisionmaking for freight transport mode and route selection are affected by these issues. For example, perishable goods and computers have high value of time (e.g. their value reduces significantly with the time.) while sand and gravel do not. Another example is that bulk and containerized commodities have different preference and, therefore, requirements for transportation means. Therefore, in order to incorporate Identify logistical criteria and classify the commodities into groups Specification of the discrete choice model’s structure and its explanatory variables Implement the proposed discrete choice model and algorithm on a simplified multimodal network Figure 4.1 Flow Chart of the Thesis Tasks Investigation of the impacts to the conventional mathematical program for network flow assignment and proposal of an algorithm for stochastic network flow assignment 40 this idea into freight transportation modeling, commodity groups should be classified according to their logistical characteristics and requirements. In this section, logistical characteristics and transportation requirements are identified. Then, Standard Classification of Transported Goods (SCTG) developed by the U.S. Department of Transportation, U. S. Bureau of the Census, Statistics Canada, and Transport Canada, which is widely used for freight transport modeling in the U.S. is reviewed and reclassified by using the logistical characteristics and requirements as the criteria. First, logistical characteristics and requirements are reviewed based on available surveys and studies. Then characteristics and requirements important to mode and route selection processes are selected and used as criteria for the classification of commodities. The 4 digit level commodity groups (283 groups) classified by STCG are used as the basis for reclassification. Clustering techniques aiming to maximize relations among commodities in the same group or mathematical programming approach aiming to minimize dissimilarity of commodities in the same group can be applied for systematic classification of commodity groups. 4.1.2 Specification of the Model's Structure and Its Explanatory Variables Basically, each commodity group classified in section 4.1 should have its own set of model parameters based on its own logistical characteristics and requirements. The potential explanatory variables to be included in the discrete choice model can be considered from two aspects: theories of supply chain and logistics and practical perspectives from shippers and carriers interviews. On one hand, from the theoretical aspect, supply chain and logistics theories and studies as well as operational models at the firm level are reviewed. Previous interviews of shippers and carriers to determine 41 influential factors in freight transport mode and route selection are also studied to gain practical insights. Based on the theoretical and practical perspectives, the important explanatory variables to be included in the discrete choice model are identified. The relationship among these explanatory variables are also identified and used for specifying the structure of the discrete choice model. For the sake of simplicity, the discrete choice model structure is assumed to be linear in its parameters (e.g. U ( 1 2 3 L q ,q ,q ,...,q ) = l l q q ... q 0 1 1 2 2 + + + + ) 4.1.3 Proposal of An Algorithm for Network Flow Assignment Based on the Proposed Discrete Choice Model According to the proposed discrete choice model for freight transport mode and route selection, the impacts and changes to the conventional mathematical program for network flow assignment are investigated. An algorithm for network flow assignment based on the proposed discrete choice model is presented. 4.1.4 Implementation of the Discrete Choice Model and Algorithm on a Simplified Multimodal Network The discrete choice model developed and the assignment algorithm are tested on a simplified multimodal network. The multimodal network comprised of the following basic elements. Freight Analysis Framework Version 2.2 highway network (Federal Highway Administration 2006) Railway network (Federal Railroad Administration and Bureau of Transportation Statistics 2006) 42 Waterway network (Vanderbilt Engineering Center for Transportation Operations and Research and Bureau of Transportation Statistics 2006) Intermodal terminals database (Oak Ridge National Laboratory and Bureau of Transportation Statistics 1998) Each individual network will be integrated via intermodal terminals to create a multimodal network. In the proofofconcept implementation, multiple origindestination (OD) pairs and pair multiple commodities are also considered. 43 CHAPTER 5 CRITERIA AND MODELS TO CLASSIFY COMMODITIES INTO LOGISTICAL FAMILIES Conventionally, the classification of commodity groups for freight transport modeling is based on industrial sectors, for example, live animals and fish (SCTG 01) and alcoholic beverages (SCTG 08) (See Appendix A 2.2). However, commodities in the same industrial sector may have different characteristics and logistical requirements. For instance, fresh food and prepared food, both of them are in the food industrial sector and classified into the same SCTG commodity group. However, their characteristics and transportation requirements are very different. Fresh food needs rapid transport mode because it is perishable while prepared food does not have such a requirement. Based on this idea, to increase the accuracy and forecasting capability of freight transport models, the commodity classification should be reconsidered based on their commodity characteristics and logistical requirements. 5.1 THE CONCEPT OF LOGISTICAL FAMILIES Many studies indicated that product characteristics affect the choice of transportation modes and routes of shippers and carriers. Abkowitz et al. (1992) studied 44 the criteria for designating hazardous materials highway routes and found criteria such as accident likelihood and population exposure/risk which are different from routing criteria of other commodities. Pedersen and Gray (1998) indicated in their study that the use of transport modes is clearly affected by the type of product transported. Jiang et al. (1999) described freight demand characteristics as a function of a firm’s characteristics, goods’ physical attributes, and the spatial and flow characteristics of shipments. Cullinane and Toy (2000) also proposed that characteristics of transported goods such as value/weight ratio and density are an important factor affecting mode/route choice. Jose HolguinVeras (2002) studied the commercial vehicle choice process and found that the shipmentsize selection is related to types of commodities. Tatineni and Demetsky (2005) also indicated that value of the commodity, density of a commodity and shelf life of goods are important variables influencing transportation decision process of a firm. The concept of commodity classification based on logistical requirements and characteristics is quite new in the context of freight transport modeling. From the literature review, there are only five studies incorporating this concept into their research. 5.1.1 Strategic Model for Integrated Logistics and Evaluations (SMILE) – the logistical requirements and characteristics used in the model are shown in Table 5.1. Table 5.1: Logistical Characteristics and Requirements in SMILE Model 1) Value/density ratio ($/kg/m3) 2) Packaging density (piece or weight per m3) 3) Value of time (low, medium, high) 4) Delivery strategy (Use or not use DC) 5) Shipment size (small, medium, large) 6) Demand frequency (frequent/ not frequent) 7) Bulk/general cargo 8) Lead time value (low, medium, high) 9) Weight/Volume ratio (kg/m3) In the model, three main characteristics e.g. value/density ratio, packaging density, volume/weight are used to classify commodities into logistical families as shown in Figure 5.1. Figure 5.1 Methodology for Commodity Classification in (Source: Based on Strategic Model for Integrated Logistics and Evaluations by Tavasszy et al., 1998) According to the logistical requirements and characteristics and classification methodology used in the SMILE model, 5.1.2 Swedish National Model System for Freight Transport (SAMGODS) SAMGODS, the logistical density (heavy or light) and value per weight unit (low or high) with other product characteristics such as dry/liquid and consumption/intermediate goods requirements and characteristics Table 5.2. Then commodities are classified into these 12 logistical families ba logistical characteristics and requirements. 45 the SMILE Model 542 types of products are sorted into 50 logistical f characteristics and requirements are bulk/general cargo, are used to develop 12 different groups as illustrated in . families. – In goods. The logistical based on their 46 Freight flow NST/R 2digits Handling category Cereals and agricultural products 00 01 04 05 06 09 17 18 General cargo Consumer food 01 11 12 13 16 Unitised Conditioned food 03 14 Unitised Solid fuels and ore 21 22 23 41 45 46 Solid bulk Petroleum products 31 33 34 Liquid bulk Metal products 51 52 53 54 55 56 General cargo Cement and manufact. build. mat. 64 69 Unitised Crude building materials 61 62 63 65 Solid bulk Basic chemicals 81 83 Solid bulk Fertiliser, plastic and other chem. 71 72 82 84 89 General cargo Large machinery 91 92 939 General cargo Small machinery 931 Unitised Misc. manufacturer articles 94 95 96 97 99 Unitised Table 5.2: Logistical Families in SAMGODS Model 1) Dry and heavy bulk goods, low value 2) Liquid and voluminous bulk goods, low value 3) Investment goods, durable consumer goods, high value 4) Heavy intermediate and consumption goods, low value 5) Lightweight consumer goods with high value 6) Lightweight intermediate and consumer goods with low value 7) Containerized bulk goods 8) Iron ore from Northern Sweden 9) High value container goods 10) Low value container goods 11) Transit goods 12) Air freight 5.1.3 SCENES European Transport Scenarios – Four main handling requirements are considered in this study: unitized, solid bulk, liquid bulk and general cargo. Commodities from Standard goods classification for transport statistics (NST/R) are grouped into 13 logistical families based on the handling requirements as shown in Table 5.3. Table 5.3: Logistical Families in SAMGODS Model 47 5.1.4 Strategic European Multimodal Modelling (STEMM) – The logistical characteristics and requirements used to classify commodities into logistical families are shown in Table 5.4. Commodities from Standard International Trade Classification (SITC) are sorted based on the criteria into 12 logistical families. Table 5.4: Logistical Characteristics and Requirements in STEMM 1) Price (low, medium, high, very high) in EURO/tonne 2) Delivery size (small, medium, large) 3) Density (low, medium, high) in tonne/m3 4) Type of goods (bulk, chemicals, parcelled) 5) Temperature control (yes, no) 6) Risk of damage (low, medium, high) 7) Level of service (low, medium, high). 5.1.5 Relations between demand for freight transport and industrial effects (REDEFINE) – Three main logistical characteristics and requirements are considered in this project: bulk/general cargo, density (high or low) and value (high or low). 14 main commodities are grouped into logistical families according to the criteria. The commodities are agricultural products, beverages and food, wood and paper, building materials, textiles and clothes, other crude minerals, chemicals and fertilizers, petrol and petroleum products, Cola and Coke, metals, machinery, transport equipment, other manufactured articles and miscellaneous articles. The logistical families are shown in Table 5.5. 48 Table 5.5: Logistical Families in REDEFINE Logistical cluster Type of goods Density (ton/m 3 ) Value (Euro/m 3 ) 1 Bulk High Low 2 Bulk Low Low 3 General cargo High High 4 General cargo High Low 5 General cargo Low High 6 General cargo Low Low High density > = 1 (ton/m 3 ) Low density < = 1 (ton/m 3 ) High value > = 5 (Euro/m 3 ) Low value > = 5 (Euro/m 3 ) Based on the review of published research and other related studies, the important logistical characteristics and requirements identified for use in this research are shown in Table 5.6 Table 5.6: Main Logistical Characteristics and Requirements in REDEFINE No. Logistical requirements 1 Value/density ratio ($/kg/m3) 2 Value of time (low, medium, high) 3 Bulk 5 Unitized 6 Weight/Volume ratio (kg/m3) 7 Dry 8 Liquid 9 Temperature control (yes, no) 10 Risk of damage (low, medium, high) 11 Harzardous material (yes, no) 12 Live (yes,no) 13 Holding cost ($/kg/m3 or piece/day) 14 Transshipment cost ($/kg/m3 or piece) It can be seen that the methodologies applied to classify commodities into logistical families is quite subjective and somewhat unsystematic. The classification highly depends on the opinions of researchers in each study. A more systematic method could be applied and results of classification improved. One possible way is to formulate the problem as a mathematical model similar to those in network clustering problems. 49 Clustering can be defined as the process of grouping objects into sets called clusters, so that each cluster consists of elements that are similar in some way. The similarity/dissimilarity criterion can be defined in several different ways, depending on the specific application and the objectives that the clustering aims to achieve. For example, in distancebased clustering two or more elements belong to the same cluster if they are close with respect to a given distance metric. On the other hand, in conceptual clustering, which can be traced back to Aristotle and his work on classifying plants and animals, the similarity of elements is based on descriptive concepts (Balasundaram and Butenko, 2006). 5.2 DISSIMILARITY COEFFICIENTS FOR COMMODITY CLASSIFICATION According to the concept of network clustering, the first task for applying to commodity classification is to define the similarity/dissimilarity of commodities. In this thesis, the dissimilarity coefficient of each pair of commodities is defined based on the weighted Minkowski metric as follows; k ki kj n k 1 ij k Max A A  D W − = = where Dij = Dissimilarity coefficient between commodities i and j Wk = Weight of importance of characteristic and logistical requirement k Aki = Characteristic and logistical requirement k of commodity i 50 Maxk = Maximum value of characteristic and logistical requirement k The details of logistical characteristics and requirements, their values and weights are illustrated in Table 5.7. Table 5.7: Logistical Requirements, Values and Weights to Calculate Dissimilarity Coefficient No. Logistical requirements Value Description Weight 1 Value/density ratio ($/kg/m3) 14 low, med, high 3 2 Value of time (low, medium, high) 13 low, med, high 3 3 Bulk 0,1 Yes/No 3 5 Unitized 0,1 Yes/No 3 6 Weight/Volume ratio (kg/m3) 13 low, med, high 1 7 Dry 0,1 Yes/No 1 8 Liquid 0,1 Yes/No 1 9 Temperature control (yes, no) 0,1 Yes/No 3 10 Risk of damage (low, medium, high) 13 low, med, high 1 11 Harzardous material (yes, no) 0,1 Yes/No 2 12 Live (yes,no) 0,1 Yes/No 3 13 Holding cost ($/kg/m3 or piece/day) 13 low, med, high 3 14 Transshipment cost ($/kg/m3 or piece) 13 low, med, high 3 5.3 OPTIMIZATION MODELS FOR CLASSIFICATION Clustering models can be classified by the constraints on relations between clusters and the objective function used to achieve the goal of clustering (see Balasundaram and Butenko, 2006). One can formulate two types of optimization problems according to a defined measure of cluster cohesiveness as follows; Type I: Minimize the number of clusters while ensuring that every cluster formed has cohesiveness over a prescribed threshold. Type II: Maximize the cohesiveness of each cluster formed while ensuring that the number of clusters that result is under a prescribed number K. The cohesiveness of a cluster in this context is measured using pairwise dissimilarity coefficients of commodities which are defined in section 5.2. The 51 optimization model to classify commodities into logistical families can be formulated in two ways as follows; Type I: Minimize the number of logistical families while limiting the maximum dissimilarity coefficient of a pair of commodities in the same logistical family. Type II: Minimize the cumulative/total dissimilarity coefficient among commodities in the same logistical family according to the specified number of logistical families. The detail of each optimization model are described as follows; Type I Objective function: = n k 1 Minimize yk Subject to x 1 for i 1, 2,..., n n k 1 ik = = = x y for i 1, 2,..., n k 1, 2,..., n ik k £ = = D z D for i 1,2,...,n, j 1,2,...,n and k 1,2,...,n ij ijk £ = = = z x x  1 for i 1,2,..., n, j 1,2,..., n and k 1,2,..., n ijk ik jk ³ + = = = z {0, 1}, y {0,1} and x {0, 1} for i 1,2,...,n, j 1,2,...,n and k 1,2,...,n ijk j ij Î Î Î = = = Where Dij = Dissimilarity coefficient between commodities i and j D = Maximum value of dissimilarity coefficient between commodities i and j to be allowed in the same logistical family xik = 1 if commodity i is assigned to family k; 0 otherwise yk = Logistical family k (if family k is assigned based on commodity k, yk = 1, else yk = 0) 52 zijk = Relationship among product i, product j and logistical family k (if product i and product j are in the same logistical family k, zijk = 1, else zijk = 0 Type II: Objective function: ij n i 1 n j 1 Minimize Dijx = = Subject to x 1 for i 1, 2,..., n n j 1 ij = = = y p n j 1 j = = x y for i 1, 2,..., n j 1, 2,..., n ij j £ = = xij Î {0, 1} for i = 1, 2,..., n j = 1, 2,..., n y {0, 1} for j 1, 2,..., n j Î = Where Dij = Dissimilarity coefficient between commodities i and j xij = 1 if commodity i is assigned to family j; 0 otherwise yj = 1 if family j is formed based on commodity j; 0 otherwise p = the number of logistical families (p is identified by users) n = total number of commodities to be classified Both integer programming formulations are similar to those of the pmedian problem (see Balasundaram and Butenko, 2006) and NPhard in general (see Gonzalez, 1985). However, in the case of commodity classification, the maximum number of commodities to be classified is less than 600 which is efficiently solvable with a good PC 53 and a commercial IP solver. In this research, classification of 49 commodities are formulated based on the developed models and solved by the optimization software XpressMP (http://www.dashoptimization.com/home//products/products_overview.html) on a computer server with a 3056 Mhz processor and 4 GB RAM. For the Type I formulation, it took less than 3 minutes to solve to the optimality, however, for the Type II formulation, it took more than 15 hours before reaching the optimal solution. The main reason for much longer computational time for the Type II formulation is because of the increased number of constraints when the number of commodity increases. 5.4 EXPERIMENTS AND RESULTS OF THE MODELS Both of the proposed approaches are applied to the U.S. commodity classification system, Standard Classification of Transported Commodity (SCTG) 4digit which has 283 groups of commodities. First, the classification of 8 commodities is used as a test case for these approaches. The details of logistical characteristics and requirements for all 8 commodities are illustrated in Appendix A3.1. The results are as follows; Results from Approach 1: Grouping (p = 3) Commodity_Index Description Group 1 1 Live animals and live fish Group 2 2, 5 Fresh or chilled vegetables except potatoes (Irish potatoes), Fresh, chilled, or frozen, except poultry Group 3 3,4,6,7,8 Dried vegetables, Dried fruit, Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered, Gasoline, Fuel oils 54 Grouping (p = 4) Commodity_Index Description Group 1 1 Live animals and live fish Group 2 2, 5 Fresh or chilled vegetables except potatoes (Irish potatoes), Fresh, chilled, or frozen, except poultry Group 3 3,4,6 Dried vegetables, Dried fruit, Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered, Group 4 7,8 Gasoline, Fuel oils Results from Approach 2: Product family Commodities Details 1 2 Fresh or chilled vegetables except potatoes (Irish potatoes) 5 Fresh, chilled, or frozen, except poultry 2 3 Dried vegetables 4 Dried fruit 6 Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered 3 1 Live animals and live fish 4 7 Gasoline 8 Fuel oils Based on the results of the commodity classification shown above, it can be seen that the commodities which were classified into different SCTG groups e.g. fresh/chilled vegetables and fresh/chilled meats in SCTG are now grouped together by these classification methods because they have similar characteristics and logistical requirements. The clustering approach was also applied to a larger set of commodities (49 commodities) in Table 5.8. The details of logistical characteristics and requirements for (D = 2) 55 all 49 commodities are shown in Appendix A3.2. The results of the experiments are as follows; Table 5.8: Commodity Index and Description Commodity Index Description Commodity Index Description 1 Live animals and live fish 25 Silica sands and quartz sands, for uses other than construction, and other sands 2 Wheat 26 Limestone and chalk (calcium carbonate) 3 Corn except sweet 27 Gravel and crushed stone except dolomite, slate, and limestone 4 Fresh or chilled vegetables except potatoes (Irish potatoes) 28 Nonagglomerated bituminous coal 5 Dried vegetables 29 Agglomerated coal 6 Fresh or chilled edible fruit except citrus 30 Fuel oils 7 Dried fruit 31 Pharmaceutical products 8 Freshcut flowers 32 Paints and varnishes 9 Unmanufactured tobacco 33 Soap, organic surfaceactive agents, cleaning preparations, polishes and creams, and scouring preparations 10 Cereal straw or husks and forage products 34 Sparkignition reciprocating internalcombustion engines for motor vehicles, of a cylinder capacity exceeding 1000 cc 11 Dog or cat food put up for retail sale 35 Parts of internalcombustion piston engines 12 Fresh, chilled, or frozen, except poultry 36 Pumps for liquids 13 Meat, salted, in brine, dried, or smoked, edible flours and meals, and pig and poultry fat, not rendered 37 Airconditioning equipment 14 Wheat flour, groats, and meal 38 Refrigerating or freezing equipment 15 Baked snack foods 39 Electric motors, generators, generating sets, and rotary converters 16 Frozen baked products 40 Electric or electronic transformers, static converters including rectifiers, and inductors 17 Milk and cream 41 Telephone or telegraph switching apparatus except parts 18 Ice cream or ice milk and their novelties, water ices, and sherbets 42 Electronic entertainment products except parts 19 Frozen vegetables and vegetable preparations 43 Computer equipment 20 Processed or prepared vegetables except frozen, dried, or milled 44 Office equipment 21 Coffee, tea, and spices, except unprocessed coffee and unfermented tea 45 Prepared unrecorded media for audio, video, computer, or other uses 22 Sweetened or flavoured water 46 Prerecorded media 23 Cigarettes 47 Electronic components 24 Silica sands and quartz sands, for construction use 48 Photographic cameras, image projectors, enlargers and reducers, projection screens, negatoscopes, and apparatus and equipment for film developing 49 Industrial processcontrol instruments 56 Results from approach 1: p = 5 Group Commodity Index 1 4,6,8,12,16,17,18,19 2 1,5,7,11,13,15,20,21,22,23,31,32 3 33,2,3,9,10,14,30 4 24,25,26,27,28,29 5 34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49 p = 8 Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 5,7,11,13,15,20,21,22,23,31 4 2,3,9,10,14 5 24,25,26,27,28,29 6 30 7 32,33 8 33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49 p = 11 Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 2,9,10,14 4 22,32,33 5 31 6 41,42,43,44,49 7 24,25,26,27,28,29 8 34,35,36,37,38,39,40,45,46,47,48 9 30 10 3 11 5,7,11,13,15,20,21,23 57 Results from approach 2: Group Commodity Index 1 1 2 4,6,8,12,16,17,18,19 3 2,9,10,14 4 22,32,33 5 31 6 41,42,43,44,49 7 24,25,26,27,28,29 8 34,35,36,37,38,39,40,45,46,47,48 9 30 10 3 11 5,7,11,13,15,20,21,23 According to the results of the classification of 49 commodities, commodities with similar characteristics and logistical requirements are grouped together. For example, in case that p = 11 for Type I formulation and D = 2 for Type II formulation, wheat, unmanufactured tobacco, cereal straw and wheat flour are grouped together. (D = 2) 58 CHAPTER 6 IDENTIFICATION OF EXPLANATORY VARIABLES AND THE STRUCTURE OF DISCRETE CHOICE MODEL In this chapter, the main purpose is to examine how logistics and supply chain management practices at the firmlevel relate to freight transport mode/route selection decisionmaking and how this concept can be incorporated into a macrolevel freight movement model. The consideration can be explored from two aspects: theories of supply chain and logistics perspective and practical perspective from shippers and carriers interview. At the end, based on both perspectives, the influential variables which should be included in the transport mode/route selection model and the structure of the model are identified. 6.1 VARIABLES BASED ON SUPPLY CHAIN AND LOGISTICS THEORETICAL PERSPECTIVE The very first thing that can be observed is the high level of interrelation between transportation decision and logistics/supply chain management at a firm level. Tatineni and Demetsky (2005) indicated that individual firms take transportation decisions as a part of the larger process of optimizing the total supply chain performance or minimizing the total supply chain costs rather than minimizing only the transportation costs. 59 SimchiLevi et al. (2003) described the relation among shipment sizes, inventory holding costs and ordering costs including transportation costs as exhibited in Figure 6.1. Figure 6.1 Relations among Shipment Size, Inventory Costs and Transport Costs Source: Designing and Managing the Supply Chain by SimchiLevi et al. (2003) From Figure 6.1, it can be obviously seen that transport choices are closely related to shipment size and inventory holding cost. More precisely, the figure shows the tradeoff between inventory cost and transport cost as shipment size varies. If one chooses to transport goods with a small shipment size, one will pay more for direct transport cost but less for inventory cost and vice versa. According to this close relation between transportation and shipmentsize decisions, many researchers proposed that the decision for transport mode and shipment size should be considered simultaneously in freight transport model. Baumol and Vinod (1970) applied inventory theory for freight transport demand model. According to their model, total cost of transport and inventory includes direct shipping costs, intransit carrying costs, ordering costs, inventory carrying costs and costs of safety stock. In the model, unit Q* Order Quantity Q Total Cost Holding Cost Ordering Cost 60 shipping cost, transit time and unit inventory carrying cost are characterization of a mode of transport. Ordering costs and inventory carrying costs are the function of shipment frequency or shipment size. Therefore, their model explicitly included the tradeoff between transport mode and shipment size decisions. Das (1974) argued that the selection of transportation services in a firm should not consider only the direct transport cost but should consider the total cost of logistics including total direct shipping cost, intransit carrying cost and the cost of consignee’s inventory operations. Constable and Whybark (1978) developed a mathematical model to minimize total logistics costs at a firm level including transportation cost, intransit inventory cost, order cost, inventory carrying cost and back order cost. The decision variables in their model can be divided into two groups: characteristics of transport modes (e.g. transport cost and transit time) and inventory variables (e.g. reorder point and order quantity). Chiang et al. (1981) proposed freight demand model which can simultaneously forecast transport mode and shipment size. The alternatives in their model are the combination of transport modes and shipments. The utility is the function of total logistics costs comprising of transport rate, capital carrying cost, intransit carrying cost, order cost and loss of value during transit and storage cost. Abdelwahab and Sargious (1992) proposed twostage conditional discrete choice model for shipment size and transport mode. First, shipment sizes of rail and road transport mode are calculated and then the calculated shipment sizes of these two modes are used as the explanatory variables to calculate transport mode choice. The attributes used in the models as explanatory variables are commodity attributes (e.g. value and density), modal attributes (e.g. freight charges, reliability and transit time) and other attributes such as regional variables, yearly demand. 61 Ostlund et al. (2002) proposed the discrete choice model for transport mode selection. In the model, total logistics costs include warehousing costs, safetystock costs at warehouse, cyclestock costs at warehouse, intransit inventory costs, trunking and delivery costs. Transport related costs (e.g. trunking and delivery costs) are the function of shipment size and this makes transport cost function nonlinear. Therefore, they proposed a linear approximation of the transport cost function by dividing shipment size into three classes: less than truck load (LTL), full truck load (FTL) and train wagon load (WL). Federal Railroad Administration (2005) developed the intermodal transportation model to study utilization between highway and railway. It was indicated that the choice of shipment sizes can affect the choice of transport modes. De Jong and BenAkiva (2007) proposed a microsimulation model for shipment size and transport chain selection. The optimal shipment size, in this research, is the one which minimizes the total logistics cost including order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs. The optimal shipment size is then used to calculate the transport mode which minimizes the total logistics cost. An important assumption for this method is that transport costs do not matter in the determination of shipment size. Based on all of the above mentioned studies, it is obvious that when a firm makes decisions for transportation services, direct transportation costs are not the only factors considered. Logistical cost components such as ordering costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs are 62 considered simultaneously. Therefore, in the context of freight modeling, these logistics costs should be incorporated into the model to better represent the decisionmaking processes of shippers and carriers. Another noticeable point is that shipment size selection largely affects the transport mode selection. 6.2 VARIABLES BASED ON PRACTICAL SHIPPERS AND CARRIERS PERSPECTIVES In section 6.1, the theoretical aspects of supply chain and logistics management have been explored. In this section, we will examine potential variables which can influence transport mode and route decisions from the practical perspectives of shippers and carriers according to conducted surveys and interviews. The details are as follows. Bardi (1973) examined the carrier selection criteria for household goods manufacturers. The five most important considerations are as follows; Reliability (meeting estimated pickup and delivery dates) Security (frequency of damage, ease of claim settlement, extent of damage) User satisfaction (courtesy of carrier employee, employee complaints, carrier reputation) Availability (carrier representative, nationwide operating authority) Transit time and cost Greeno et al. (1977) investigated the transport attributes affecting transport mode selection comprising of rail freight, rail express, private fleet, steamship, truck forwarder and common carrier. The investigated list of attributes is as follows;  Ontime delivery  Time in transit  Expensiveness  Frequency of service  Tracing time  Completeness of service 63  Promptness of claim and settlement  Competence of staff  Equipment availability  Flexibility of service  Simplicity of dealing  Intermodal flexibility  Care with shipments  Innovativeness  Quality of personal According to the interview and questionnaire of 80 firms, it was found that six attributes were the most important and almost totally explain the decisions. These six attributes are time in transit, inexpensiveness, frequency of service, tracing time, flexibility of service and care with shipments. McGinnis (1980) applied the factor analytic approach to examine factors influencing shippers’ choice of transportation modes. The results ranked in order of importance can be shown as follows; Speed and reliability Freight rates Loss and damage External market influences Inventories Market competitiveness Company policy and customer influences McGinnis (1989) investigated the results of 11 studies related to transportation mode choice. He concluded that the factors affecting transport mode selection could be classified as follows; Freight rates – costs, charges, rates Reliability – reliability, delivery time 64 Transit time – timeintransit, speed, delivery time Over, short and damaged  loss, damage, claim processing and tracing Shipper market considerations – customer service, user satisfaction, market competitiveness, market influence Carrier considerations – availability, capability, reputation, special equipment Product characteristics – product perishability, packaging requirements, new products Jeffs and Hills (1990) studied the determinants of transport mode choice of shippers in the paper, printing and publishing industry in Britain. They found that the determinants are reliability, control over dispatch and delivery time, avoidance of damage to goods, security of product in transit, transit time, ready to transport when required, length of haul and size of consignment. Abkowitz et al. (1992) studied the criteria for highway route choices of hazardous materials transportation. They proposed that, at least five criteria should be considered for route selection to make the decision safe for public and efficient simultaneously. The criteria are shipment, distance, travel time, accident likelihood, and population exposure. Indeed, the multiplication of accident likelihood and population exposure is the classic definition of risk in hazardous material transportation. Matear and Gray (1993) studied the factors influencing freight service choice for shippers and freight suppliers in Irish sea freight market. Of thirty investigated factors, the most important factors included fast response to problems, avoidance of loss or damage, ontime collection and delivery, value for money price, good relationship with carrier, ability to perform unanticipated urgent deliveries, short transit time, low price and ability to handle shipments with special requirements. Murphy et al. (1997) studied the transport mode selection criteria of shippers and carriers and compared the difference. The results are shown in Table 6.1. 65 Table 6.1: Mean Score of Factors Influencing Transport Mode Selection of Shippers and Carriers Shipper Carrier Reliability 1.22 1.28 Equipment availability 1.36 1.79 Transit time 1.45 1.83 Pickup and delivery service 1.60 1.93 Financial stability 1.63 2.07 Operating persennel 1.66 1.81 Loss and damage 1.69 2.11 Rates 1.71 2.57 Service frequency 1.83 2.09 Scheduling flexibility 1.85 1.96 Expediting 1.88 2.07 Rate changes 1.88 2.58 Service changes 1.89 2.05 Tracing 1.91 2.63 Linehaul services 1.98 2.33 Claims 2.05 2.72 Carrier salesmanship 2.68 2.89 Special equipment 3.05 2.51 Mean score Factor Source: Carrier Selection: Do Shippers and Carriers Agree, Or not by Murphy et al. (1997) As shown in Table 6.1, the relative importance of influential attributes is quite the same between shippers and carriers. The most important factors are reliability, equipment availability, operating personnel, transit time, pickup and delivery service and financial stability. Pedersen and Gray (1998) studied the transport selection criteria of shippers in Norway. The investigated determinants included Timing factors – reliability in collection and delivery time, high transport frequency, short transit time and directness of the transport route Pricing factors – freight rate, difference between actual and estimated costs, special offer/discount, packing charges Security factors – damage/loss frequency, control over delivery time, ability to 66 monitor the goods in transit, knowledge of port/labor Service factors – coordination and cooperation with carriers, flexibility, ability to handle urgent deliveries and ability to handle special consignments They concluded that, for Norwegian shippers, transport price factors are the most important ones. However, the influential factors highly depend on product characteristics. For example, the exporters with high valuetoweight ratio consider the timing factors more important than pricing factors. Jiang et al. (1999) studied the demand characteristics affecting transport mode selection from revealed preference data in France. They classified demand characteristics into three types as follows; A firm’s characteristics include type of firm (for example, factories, shopping centers, or warehouses), the firm’s structure (small, nationwide or worldwide) and the firm’s location (for example, the accessibility to rail branch lines and highways). A firm’s own transportation facilities closely relate to its transportation demand and are also an important factor in its modal choice. In addition, a firm’s information system strongly influences its logistic practices and plays an increasingly important role in its transportation decisions. Another demand characteristic is attributes of the goods to be transported, such as type of product, weight, value, and packaging. Packaging is generally either parcels and pallets, containers, and cases. The last demand characteristic is spatial distribution and physical flow including frequency, distance and origin/destination of shipments. It was found that transportation distance, company size and type, information system, accessibility to transport infrastructure from origin and destination, shipment packaging, 67 truck ownership, shipment size are critical determinants for transport mode decision. Cullinane and Toy (2000) reviewed 75 articles related to freight transport mode/route selection criteria and applied the content analysis to identify critical attributes that have been most mentioned in the literature. The list of the attributes mentioned in the reviewed literature is shown in Table 6.2. According to the results of the content analysis, the most five important factors are cost/price/rate, speed, transit time reliability, characteristics of goods and service (unspecified). Norojono and Young (2001) investigate the shippers’ perception of rail freight services in Indonesia. In this research, the investigated attributes are as follows; Transport charge Delivery time Service quality represented by delivery time reliability measured by the probability of being late due to irregular service, safety with respect to the number of cargoes having loss or damage, distance to rail terminal, and train type whether it is freight train or part of passenger train Flexibility of the service represented by frequency measured by the number of service in each day, departure time, responsiveness measured by complaints The results of the study showed that reliability of delivery time and loss/damage are the major influences. In addition, it is reasonable to define flexibility as the function of frequency and responsiveness. 68 Table 6.2: Category of Attributes and Underlying Terms Category Name Terms covered by category Cost/Price/Rate Cost, price, rate Service (nonspecified) Service (nonspecified) Transit time reliability Transit time reliability Frequency Frequency Distance Distance Speed Speed, transit time, terminal time, transshipment time Flexibility Flexibility, convenient schedule, nonspecific extras, pickup and delivery Infrastructure availability Infrastructure availability, accessibility Capability Capability, service availability, equipment availability, capacity Inventory Inventory Loss/Damage Loss/Damage, claims Characteristics of the goods Type, value, value/weight ratio, volume, weight, density, shipment size Sales per year Sales per year Controllability/Tracability Controllability, tracability Previous experience Nonspecific positive behavior, relationship, image of modes used, stability of firm Source: Identifying Influential Attributes in Freight Route/Mode Choice Decisions: A Content Analysis by Cullinane and Toy (2000) Jose HolguinVeras (2002) studied the choice of shipment size and types of vehicle. It is indicated that shipment size could be adequately modeled as a function of the trip distance, the type of commodity being transported, and the type of economic activity taking place at the origin and destination of the trip. The economic activities at origin and destination, in this case, are retail, wholesale and other activities. Tuna and Silan (2002) investigate the selection criteria for transportation services of liner shippers in Turkey. Table 6.3 shows the relative importance of each criterion. 69 Table 6.3: Relative Importance of Transport Selection Criterion Overall Standard Mean Score Deviation Delivering the cargo without damage 4.84 0.44 Issuing accurate shipping documentation 4.81 0.47 Delivering the cargo at the promised time 4.68 0.63 Dependability in handling problems 4.68 0.53 Informing of changes to schedules 4.62 0.55 Issuing accurate price quotations 4.54 0.77 Responding to complaints quickly 4.54 0.77 Issuing shipping documentation quickly 4.51 0.77 Responding to urgent deliveries quickly 4.49 0.93 Giving clear & correct information about costs 4.43 0.87 Issuing accurate invoices 4.43 0.87 Transit time 4.38 0.86 Willingness of the personnel to help 4.35 0.72 Responding to enquiries promptly 4.35 0.79 Minimum changes to schedules 4.27 0.9 Providing clean& undamaged equipment 4.27 0.87 Expert and knowledgeable personnel 4.19 0.81 Informing whether goods will be transshipped 4.16 0.83 Polite and Respectful personnel 4.11 0.77 Informing about the condition of the cargo 4 0.91 Giving arrival notices on time 4 1.18 Convenient working hours for contact 3.89 0.84 Issuing invoices on time 3.73 1.02 Providing special equipment 3.62 0.92 STATEMENTS Source: Freight Transportation Selection Criteria: An Empirical Investigation of Turkish Liner Shipping by Tuna and Silan (2002) As seen from Table 6.3, ‘Delivering the cargo without damage’, ‘issuing accurate shipping documentation’, ‘delivering the cargo at the promised time’ and ‘dependability in handling problems’ were determined as the most important factors. Mangan et al. (2002) studied the attributes affecting the selection of port and ferry services in Ireland. In their study, there are 14 important attributes affecting the choice of port and ferry services ranked in order of importance as shown in Table 6.4. 70 Table 6.4: Influential Attributes for Port/Ferry Choice (Ranked in Order) 1. Space available when needed on ferry 2. Sailing frequency/convenient sailing times 3. Risk of cancellation/delay 4. Port and ferry on fastest overall route 5. Proximity of ports to origin/destination 6. Cost of ferry service/discounts 7. Speed of getting to/through ports 8. Port/ferry on cheapest overall route 9. Ferry suitable for special cargo 10. Delays due to driving bans, etc. 11. Availability of info on sailing options 12. Facilities for drivers 13. Opportunity for driver rest break 14. Intermodal/connecting transport links Rank of Important Attributes for Port/Ferry Choice Source: Modelling Port/Ferry Choice in RoRo Freight Transportation by Mangan et al. (2002) MorenoQuintero and Watling (2002) studied the route choice behaviors of truck drivers in Mexico and found that variables affecting such a decision were direct travel cost, fines for overloading, enforcement strategies, level of service and congestion. Beuthe et al. (2003) studied attributes influencing transport mode selection by using stated preference survey. The commodities were steel, textile, electronics, chemical, cement, packing and pharmacy. The transport modes included rail, road, waterway, shortsea shipping and intermodal transport. The attributes expected to be influential are as follows; Cost, i.e. outofpocket cost for transport, including loading and unloading; Time, i.e. doortodoor transport time, including loading and unloading; Loss as the % of commercial value lost from damages, stealing and accidents; Frequency of service per week proposed by the carrier or the forwarder; Reliability as the % of deliveries at the scheduled time; Flexibility as the % of times nonprogrammed shipments are executed without undue delay. 71 The results from their analysis showed that the importance of each attribute is different for the different commodities. For instance, time and reliability are important for the textile firm and the producer of electronics, which ship over rather long distances. Reliability, flexibility and losses appear important for the pharmaceutical firm, which seems ready to pay for it. Vannieuwenhuyse et al. (2003) conducted an online survey of shippers and logistics service providers in Belgium to collect data about attributes affecting selection of transport modes. Five most important attributes were identified: transportation cost, reliability, flexibility, transportation time and safety. Punakivi and Hinkka (2006) investigated the selection criteria for transportation modes of shippers in four Finnish industry including electronics, pharmaceutical, machinery and construction. They concluded that the selection criteria for transportation modes are different depending on product characteristics in each industrial sector. The selection criteria for each industry ranked in order of importance are presented in Table 6.5. Table 6.5: Selection Criteria for Transportation Modes Rank of Importance Electronics Pharmaceutical Machinery Construction 1 Quality Speed Price Price 2 Speed Convenience Reliability Scheduling 3 Price Safety Punctuality Punctuality 4 Convenience Fluency Speed Convenience Source: Selection Criteria of Transportation Model: A Case Study in Four Finnish Industry Sectors by Punakivi and Hinkka (2006) Note: quality covers reliability, accuracy and safety. Convenience represents the ability to take special product characteristics into account in operations. Based on the above studies, it is evident that there are some factors besides transportation and logistics costs influencing transport mode and route decisions. However, at the macro level freight demand model, the factors which should be 72 considered are shown in Table 6.6. Table 6.6: Main Influential Factors at Macro Level 1 Transit time 2 Transit time variability 3 Frequency of loss/damage 4 Ease of claim settlement 5 Frequency of service 6 Tracing capability 7 Availability of special handling equipment 8 Flexibility of service 9 Product characteristics 10 Accident likelihood 11 Population exposure 12 Accessibility to transport infrastructure 13 Economic activities at origin/destination 14 Enforcement on highway route Factors Affecting Transport Mode/Route 6.3 RELATIONSHIP AMONG ATTRIBUTES In this section, the relationship among all factors described in section 6.1 and 6.2 are considered. From the review in section 6.1 and 6.2, it can be concluded that freight transportation mode/route decisions can be affected by the factors as follows; Shipment size Total logistics cost including order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs Service quality attributes including transit time, transit time variability, frequency of loss/damage, ease of claim settlement, frequency of service, tracing capability, availability of special handling equipment, flexibility of service, product characteristics, accessibility to transport infrastructure, economic activities at an 73 origin/destination, enforcement on highway route, accident likelihood and population exposure (Note: the last two attributes are for hazardous materials) All of the above influential factors can be included into the model as follows; For shipment size, this factor can be incorporated into the model either as a decision variable or an explanatory variable. However, as exhibited in Figure 6.1, the transport decision is highly dependent on shipment sizes. In addition, based on the previous research experiences by Chiang et al. (1981) and De Jong and BenAkiva (2007), it has been shown that it is reasonable to construct the discrete choice model for transport mode and shipment size simultaneously. Therefore, in this thesis, shipment size will be incorporated into the discrete choice model as a decision variable. Figure 6.2 shows the interaction among all related factors e.g. logistics costs, transport mode/route characteristics, shipment size and other related factors According to Figure 6.2, shipment size interacts with all types of logistics costs e.g. order costs, transport costs, consolidation and distribution costs, costs of deterioration and damage during transit, capital costs of goods during transit, inventory costs and capital costs of inventory and stockout costs Shipment size can be affected by product characteristics and an activity at origin/destination such as wholesale, retail or manufacturing Transit time and transit variability affect transport costs, capital costs of goods in transit, inventory costs, capital costs of goods in inventory and stockout costs Frequency of loss/damage affects costs of damage during transit 74 Figure 6.2 The Interaction among Related Factors Transport Mode/Route Characteristics  Transit time  Transit time variability  Frequency of service  Ease of claim settlement  Frequency of loss/damage  Tracing capability  Availability of special handling equipment  Flexibility of service  Accident likelihood  Population exposure  Enforcement on highway route Other Factors  Product characteristics  Accessibility to transport infrastructure  Activity at origin/destination Total Logistics Costs  Order costs  Transport costs  Consolidation/distribution costs  Costs of damage during transit  Capital costs of goods in transit  Inventory costs and capital costs of inventory  Stockout costs Shipment Size 75 6.4 IDENTIFICATION OF INFLUENTIAL ATTRIBUTES AND MODEL STRUCTURE This section explains how to include all explanatory variables described in section 6.3 into the discrete choice model. The details are as follows; 6.4.1 Shipment Size – In this thesis, shipment sizes are divided into 3 intervals: 0 ton shipment sizes < 25 tons (e.g. equivalent to 1 truckload), 25 tons shipment sizes < 50 tons (e.g. equivalent to trailer truck) and 50 tons shipment sizes < 80 tons (e.g. more than a trailer truck but with in capacity of rail and ship). The combination of a shipment size and a transportation path forms an alternative in the discrete choice model. 6.4.2 Transit Time –Total transit time of a path from an origin to a destination is the summation of transit time of each link and at each transshipment node in that path as shown in Figure 6.3. For railway/waterway links and transshipment node, they are assumed to be uncapacitated and their transit times are assumed to be normally distributed with a constant mean. However, for a highway link, it is assumed to be capacitated and its transit time is dependent on link’s flow, capacity and free flow travel time in the following forms; = + H ) Q V TT FT 1 ( Where TT is transit time FT is freeflow travel time of a highway link 76 V is traffic flow on that link Q is capacity of the link and are parameter ( is 0.15 and is 4, Bureau of Public Roads, 1964) Figure 6.3 Transit Time of A Path 6.4.3 Transit Time Variability – In this thesis, transit time and transshipment time are assumed to be normally distributed and independent from those of other links. Transit time variability can be measuered by variance of transit time which can be calculated based on coefficient of variation and transit time of a link e.g 2 2 = (C.V. × mean transit time) . Total variance of transit time 



A 

B 

C 

D 

E 

F 

I 

J 

K 

L 

O 

P 

R 

S 

T 

U 

V 

W 


