A Multi-Agent Reinforcement Learning Framework with Recurrent Communication Module for Traffic Light Control

As a significant issue in the development of Smart City, traffic light control is closely related to public transport. Multi-agent reinforcement learning is a feasible paradigm to control decentralized systems. In this work, we design a recurrent communication module for traffic light control systems. Aiming at lightweight decentralized learning, we decouple the interaction among entities, using an improved recurrent neural network to share messages. Regarding traffic lights as a partially observable Markov Decision Process, we develop a complete reinforcement learning model. Every light receives messages from its neighbors through the improved recurrent neural network. For the further lights, its neighbor lights can play the role of intermediation to transmit messages. By our communication framework, traffic lights can learn their control policies in a decentralized way, and distribution control can be executed through a communication channel. Several simulated experiments show that our method outperforms the fixed time method and the existing reinforcement learning-based frameworks under different traffic demands.