Author(s)

T. Miyagi & M. Ishiguro

Abstract

The conventional models for describing car-drivers’ route choice behaviours in traffic networks have treated the decision-makers as a non-atomic quantity who are homogeneous in the preference function and always take rational behaviours. Those behavioural assumptions require that each driver knows the minimum-cost route in spite of the deterministic or probabilistic. The action hypothesis in a route choice is not realistic and is insufficient to analyze the influence that traffic information gives over action choice. In this study, we treat each driver as a discrete decision-maker and assume as a heterogeneous agent with bounded rationality. Each agent does not know the minimum-cost route on the network, and only knows the route information that he or she has experienced. This assumption motivates us to propose a behavioural model in which regretmatching is combined with reinforcement learning. We show that even in such a situation there exist adaptive learning rules that lead drivers to rational choices in the long run. Keywords: route choice behaviour, regret matching, reinforcement learning. 1 Introduction The transportation system is a complex system where the decision-making of each agent is mutually related and influenced. This paper proposes a new approach for describing route-choice behaviours of agents in transportation networks based on the theory of games. The conventional procedure for predicting flows in the transportation network has so far been constructed on the basis of the user equilibrium concept: the population of decision-makers is assumed to be homogeneous and infinitely divisible. It is also assumed that such a representative user has a set of complete travel information about his route

Keywords

route choice behaviour, regret matching, reinforcement learning.