Exploring machine learning for adaptive traffic signal control

Under conditions of severe local congestion on urban road networks, such as might be caused by major events, incidents or construction activity, existing traffic signal control and coordination methods may not generate optimum control strategies. Unless managed effectively, traffic queues can cause gridlock, bringing traffic flow to a virtual half. In this paper, we explore the application of reinforcement-learning (RL) agents to the task of controlling a traffic signal. RL-agents require neither a complete model of the environment, nor exemplary supervision. RL-agents learn from their experience as they dynamically interact with a changing environment We describe the control of an isolated traffic signal using an RL-agent Under traffic conditions of lower variability favourable to pre-timed control, the RL-agent performs on a par with or slightly better than a pre-timed signal controller. However, under more variable traffic conditions, the RL-agent demonstrates marked superiority. We also outline our ongoing research effort to extend RL-based control to a signal system and integrate this with dynamic route guidance.