DEVELOPMENT OF MAXIMUM ENTROPY ESTIMATOR FOR CALIBRATING TRIP DISTRIBUTION MODELS

• Main Author: ADI WIDIARJANA, NYOMAN
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/2184
• Institution: Institut Teknologi Bandung

With the growth of development the need for transport infrastructure is also increasing. To alleviate transportation problems good planning, supported by accurate data of travel patterns of the study areas, are needed. In developing countries where each area has a different and rapidly changing travel pattern, and with limited funds, time, and labor, it is difficult to obtain accurate patterns. Thus the development of a method which utilizes available patterns to produce accurate future patterns at relatively low cost, is needed. This study developed a method for approximating trip patterns, i.e. trip distributions, from total attraction and generations, based on the maximum entropy (ME) approach. Basically this method assumes movements as those from gas molecules moving unimpeded and dispersed. Several estimation methods have been developed earlier, i.e. the Non Linear Least Square (NLLS), Maximum Likelihood (ML) and Inference Bayes (IB) methods, and these were used as comparison for the developed estimation method. This study used the gravity model as trip distribution model with its three types of constraints (DCGR, PCGR and ACGR models), and three deterrence functions (exponential, power and Tanner's), and then compared and applied all the four estimation methods. The results of this study are: 1) the computer programme of the developed method (ME) performed well ; 2) with artificial data the most suitable one, with which the best approximations were obtained, are : for NLLS it is uniformly distributed data; for ME it is wide ranging data; 3) for the East Java study'case, for private car movements in 1991, the combination of NLLS, DCGR and Tanner's function produced the best approximation, i.e. RMSE=139.8794 and R2=0.9765, which is only slightly better than with ME; 4) the trip patterns influence the results of estimation; different patterns need different parameter values and model combinations. The study recommends to expand and test the method with other real data.