Reinforcement learning-based traffic lights controller with adaptive reward function

Traffic Congestion affects the economy, the sustainability of the urban environment and the citizens’ well-being. This problem can be alleviated by promoting more efficient usage of the road network through intelligent Traffic Control strategies. Reinforcement Learningbased Traffic Signal controllers offer many benefits in relation to other techniques, and one of them is the ability to tune the actions of the controller by customising its reward function. In this paper, we evaluate how different reward functions behave under different demand conditions and propose an Adaptive Reward Function that dynamically adapts its goal according to the road’s saturation levels