Sensitivity of reinforcement learning agents to aggregated sensor data in congested traffic networks

Flexible signal timing operation with cycle-free and sequence-free strategies using reinforcement learning has been researched from different fields and applied to transportation networks. Such techniques naturally rely on accurate incoming data for optimal operation. However, the effect of imperfect information received by RL agents in a traffic environment has not been explored in detail and may provide further indication to whether they can be truly suitable for real-world applications. This paper studies this topic in the context of a congested traffic network, where RL agents receive aggregated loop detector data to make decisions, instead of directly observing activations from all vehicles. A case study shows the sensitivity of the agents' performance when data is aggregated to different levels. Aggregation levels are used as a method to represent imperfect information, and the performance of the system is used as an indicator to determine acceptable aggregation for the system to remain operational in oversaturated conditions.