Multi-Objective Traffic Signal Control Using Network-Wide Agent Coordinated Reinforcement Learning

Abstract

The optimization of intersection signal control can improve traffic efficiency, reduce congestion degree, and improve traffic safety. Aiming at implementing the coordinated adaptive traffic signal control (ATSC) across large-scale arterial network, multi-agent reinforcement learning (MARL) has been widely concerned and lucubrated. Nevertheless, the existing MARL-based ATSC studies suffers from several limitations: (1) While most existing researches focused on the mobility performance of controlled corridor, there calls for a methodology that aims at combine multi-objective performance on traffic safety, efficiency, and network coordination simultaneously; (2) Most methods ignore the correlations between multiple agents, nor considers the spatial-temporal dependencies among the corelated neighboring intersections due to high communications requirements, which can hardly be achieved in real adaptive coordination control. To overcome the aforementioned difficulties, a multi-objective reinforcement learning model (NACRL) for network-wide coordinated signal control is proposed. Firstly, to enforce a coordinated network control with safety and efficiency considerations, a reward mechanism inspecting both traffic safety and traffic efficiency indicators was designed to achieve ideal performance in terms of mobility, safety and smooth. Secondly, the proposed NACRL conducted a centralized training-decentralized execution framework, this overcomes the critical limitation of data transmission in the field implementation while explicitly analyzing the traffic state over the entire network instead of examining each isolated intersection. Last but not least, the proposed algorithm utilized the attention mechanism to dynamically capture the sophisticated spatial-temporal dependencies over the complex arterial network, which aids the better coordinated control over multi-agents deployed at the intersections across the corridor. To testify the effeteness of the proposed algorithm, extensive experiments were implemented in both large-scale synthetic traffic grid and real-world arterial network. The experiment demonstrated that the proposed NACRL algorithm outperforms other state-of-the-art baselines with simultaneously improved performance in terms of traffic safety, traffic efficiency and network coordination, as well as improved algorithm convergence and interpretability.

Keywords: Multi-objective reinforcement learning, Network-wide adaptive traffic signal control, Attention mechanism, cooperation and coordination