DESIGN OF AN ARTIFICIAL SIMULATOR FOR ANALYZING ROUTE CHOICE BEHAVIORIN THE PRESENCE OF INFORMATION-SYSTEM

Computer simulation is often-used methodology to study travel behavior as a cost effective alternative to field studies. In this study, we u tilize PC-based computer simulation to study the effects of informatio n on route choice and learning. Building on the efforts of a prior sta ge of simulation, further experiments that utilize an expanded traffic network and provide various levels of information to subjects, have b een designed. This framework allows us to investigate both pretrip and en route route-choice behavior, and capture the effect of different l evels of information of drivers' learning and adaptive processes that are being undertaken in these experiments. The experiments were design ed in two stages. In the first stage, a simple, two route-alternative traffic network was developed. Experiments conducted with this network provided extensive comments from participants, which were modeled usi ng object-oriented programming techniques to produce a better subseque nt design. The data from the first stage was analyzed using neural net work techniques and the network was trained using back-propagation. Th e second stage of experiments utilized a multiple-route, expanded netw ork with pretrip and/or en route information, and varying levels of in formation. The data obtained in this stage is being analyzed using rec urrent neural networks. This paper describes the design and analysis o f the first stage of experiments, and the redesign of the network simu lation using experience gained in the first stage. The design of the n etwork simulation involved the following steps: requirements analysis, database design, specifications of user-computer interface, design of shortest path module, software develop ment, and prototype testing an d refinement. The simulator was developed using an object-oriented pro gramming language, C++. The object-oriented features included inherita nce, class hierarchy, message passing and concurrence. A recurrent neu ral network has been built for future modeling of the data obtained in the second stage of experiments. This neural network will be used to predict subjects' choices of whether or not to follow the system-provi ded advice, depending pn past experience. An important feature of the neural network is that the decisions at previous nodes, will be used a s an input at subsequent nodes. This allows us to model participants r oute choice behavior at every node, that is at every decision point.