Dynamic Traffic Signal Control for Isolated Intersections: Enhanced SARSA Reinforcement Learning with Expectation Prediction and Eligibility Traces.

With the development of intelligent transportation systems, new methods emerge to overcome the limitations of traditional fixed-timing traffic signal control. Among these methods, intelligent signal timing control based on reinforcement learning (RL) boasts promising prospects. Classic RL algorithms have flaws like high estimation bias and lack of robustness, while the newer complex-structure algorithms have issues with training time and unstable convergence speed. To overcome these limitations, this paper develops a two-branch cooperative signal control framework integrating Expected SARSA and SARSA(λ) for isolated signalized intersections, constructing a traffic-adaptive Markov Decision Process (MDP) dynamic decision model using real-time traffic flow data. Expected SARSA reduces policy variance and Q-value overestimation by calculating expected action values. SARSA(λ) adopts eligibility traces for multi-step temporal difference error backpropagation to significantly boost sample utilization efficiency and overall model robustness. Experiments cover algorithm comparison, framework generality and hyperparameter robustness. Experimental results reveal that under the framework proposed in this paper, the two improved algorithms outperform the baseline algorithms in quantitative evaluations conducted from three dimensions. Scenario and parameter tests validate its generality and robustness. This work offers an efficient and reliable solution for real-time adaptive signal control of isolated intersections.