Dealing with continuous-state reinforcement learning for intelligent control of traffic signals

Reinforcement learning is an efficient, widely used machine learning technique that performs well in problems that are characterized by a reasonable number of states and actions. This is rarely the case in multiagent learning problems. In such cases it is necessary to use techniques that generalize the state space such as function approximators. The focus of this work is to combine multiagent learning with a new neural network model, called Incremental Gaussian Mixture Network (IGMN), which is able to learn incrementally using a single scan over the training data (each training pattern can be immediately used and discarded). Thus this approach is key in scenarios where agents have a high number of states to explore. This is the case in traffic signal controllers when the state space is a continuous variable. In this scenario, our results indicate that the proposed representation outperforms the tabular one, thus being an effective alternative for learning traffic signal control policies.