A3C and A2C performance comparison in intelligent traffic signal controller using Indonesian traffic rules simulation

Traffic congestion is one of the most common problems in line with the increase in population and economic activity. To tackle this problem, many places in the world place a traffic light in most major or dense traffic intersections. The traffic lights, especially in Indonesia, are conventionally giving the signals; it always repeats the same cycle over time. This practice has a weakness in its inability to adapt to various traffic conditions. Reinforcement learning is a popular approach for intelligent traffic signal controller as it has advantages such as self learning without the need of supervision, goal oriented, real-time adaptation and curse of dimensionality management. One of the algorithms that can be used is Asynchronous Advantage Actor-Critic (A3C). This paper will discuss the experiments of Asynchronous Advantage Actor-Critic performance compared to stable baseline Advantage Actor-Critic performance using real-life-like traffic rules environments in SUMO with an open-source repository using the Indonesian traffic rules Environment. Results given in this study show that Asynchronous Advantage Actor-Critic (A3C) got a faster learning time and effective in gaining optimal result compared to Advantage Actor-Critic (A2C) using real-life-like traffic rules environments in SUMO with an open-source repository. Furthermore, our results indicate that the advantage of using A3C is not that visible in the single intersection scenario, but notable in the four and nine intersections scenario.