Multiagent Learning-Based Approach to Transit Assignment Problem

Abstract

This paper presents an operational prototype of an innovative framework for the transit assignment problem, structured in a multiagent way and inspired by a learning-based approach. The proposed framework is based on representing passengers and their learning and decision-making activities explicitly. The underlying hypothesis is that individual passengers are expected to adjust their behavior (i.e., trip choices) according to their experience with transit system performance. A hypothetical transit network, which consists of 22 routes and 194 stops, has been developed within a microsimulation platform (Paramics). A population of 3,000 passengers was generated and synthesized to model the transit assignment process in the morning peak period. Using reinforcement learning to represent passengers’ adaptation and accounting for differences in passengers’ preferences and the dynamics of the transit network, the prototype has demonstrated that the proposed approach can simultaneously predict how passengers will choose their routes and estimate the total passenger travel cost in a congested network as well as loads on different transit routes.