Adaptive Group-Based Signal Control Using Reinforcement Learning with Eligibility Traces

摘要

Group-based signal controllers are widely deployed on urban networks in the Nordic countries. However, group-based signal controls are usually implemented with rather simple timing logics, e.g. vehicle actuated timing. In addition, group-based signal control systems with pre-defined signal parameter settings show relatively poor performances in a dynamically changed traffic environment. This study, therefore, presents an adaptive group-based signal control system capable of changing control strategies with respect to non-stationary traffic demands. In this study, signal groups are formulated as individual agents. The signal group agent learns from traffic environments and makes intelligent timing decisions according to the perceived system states. Reinforcement learning with multiple-step backups is applied as the learning algorithm. Agents on-line update their knowledge based on a sequence of states during the learning process rather than purely on the basis of single previous state. The proposed signal control system is integrated into a software-in-the-loop simulation (SILS) framework for evaluation purpose. In the testbed experiments, the proposed adaptive group-based control system is compared to a benchmark signal control system, the well-established group-based fixed-time control system. The simulation results demonstrate that learning-based and adaptive group-based signal control system owns its advantage in dealing with dynamic traffic environments in terms of improving traffic mobility efficiency.