Microscopic Model-Based RL Approaches for Traffic Signal Control Generalize Better than Model-Free RL Approaches

There have been many recent advances in the Traffic Signal Control literature that use reinforcement learning, most of which is undertaken using the model-free approach. Approaches in the model-free domain, attempt to learn the value or the policy function directly without attempting to learn the environment transition dynamics. Therefore, training the value function under a specified dynamics fails to differentiate the value updates from the underlying dynamics, making these methods require much larger agent-environment interaction data to generalize over different scenarios. In contrast, approaches that optimize agent actions w.r.t. a learned dynamics model inherently avoid this tight coupling of dynamics and value, allowing for much faster adaptation as traffic scenarios change. For this work on single intersection control, we specifically adopt this latter model-based approach of learning a microscopic simulator model and then apply tree-search techniques to optimize control actions. This approach quickly generalizes to a diverse set of traffic demands, whereas the model-free method performs suboptimally in conditions unseen during training. Another benefit of model-based approaches is the ability to control new intersections with previously unseen topologies, which makes the method transferable in terms of both demand and intersection structure variation. Finally, we observe that pairing these control strategies with the learned model also makes our approach debuggable and explainable, which is a critical requirement for real-world deployment.