Empirical Learning of Decision Parameters for Agent-Based Model

Agent-Based Model (ABM) is a widely used tool to analyze distributed systems. However, the decision-making parameters are difficult to determine, since ABM is a kind of micro model and such parameters, varying from person to person, cannot be measured conveniently in real traffic systems. For this problem, this paper introduces reinforcement learning to empirically and efficiently calculate the micro parameters of ABM. By a parameterization of the individual interactions, our new approach is able to decouple the dependence for a given agent upon his “social neighbors”, and thus can accelerate the learning process. Experiments on inter-city traveling of population indicate that the proposed method is effective for the micro parameter computation.