The traffic congestion becomes a severe problem in almost every city and the applications of Intelligent Transportation Systems makes it possible for an adaptive traffic signal control system to improve signal control strategy. Exploiting deep reinforcement learning for traffic signal control is a frontier topic in intelligent transportation research. However, centralized reinforcement learning is hard to be used for large-scale traffic signal control system due to the high dimensions of the joint action space. Multi-agent deep reinforcement learning overcomes the curse of dimensions but introduces a new problem: how to learn coordination between different agent under a partially observable traffic environment. In this paper, we introduce a multi-agent deep reinforcement learning algorithm for a large-scale traffic signal control system. The proposed method is compared with greedy policy, independent Q-learning method and independent actor critic method in a large synthetic traffic networks on the SUMO microscopic traffic simulation platform. The simulation demonstrate the proposed method is more efficient than other decentralized reinforcement learning approach.

CICTP