Theory of Reinforcement Learning and Algorithms of Route Choice in Transportation Networks

研究表明，将交通网络中的个体出行者严格建模为一个自适应学习智能体，该智能体通过日常经验获取出行信息，并根据已实现的收益做出决策，从而强化自己的行动。提出了一种符合理论的自适应学习算法，并证明了该算法使系统达到概率为1的纳什均衡。通过各种链路代价函数定义不清的网络实例对所提算法进行了数值测试，并检验了算法的快速收敛性。此外，我们还对路径选择模型中包含的结构参数提出了一种估计方法。对室内实验所得的日常路线选择数据的应用效果令人满意。

This research shows that an individual traveler in transportation networks is rigorously modeled as an adaptive learning agent who receives travel information through day-to-day experience and makes his decision so as to reinforce his action depending the realized payoffs. An adaptive learning algorithm consistent with the theory is proposed and proved that it leads the system to a Nash equilibrium with probability one. The proposed algorithms have tested numerically by using example networks with various ill-defined link cost functions and examined a rapid convergence of the algorithms. In addition, we have proposed an estimation method for the structure parameters included in the route choice model. The application to the data of theday-to-day route choice obtained by the indoor experiments was satisfactory.