Learning to Control Forest Fires

Forest fires are an important environmental problem. This paper describes a methodology for constructing an intelligent system which aims to support the human expert's decision making in fire control. The idea is based on first implementing a fire spread simulator and on searching for good decision policies by reinforcement learning (RL). RL algorithms optimize policies by letting the agents interact with the simulator and learn from their experiences. Finally, we observe different problems and propose solutions for solving them. Among these problems are storing policies for huge state spaces and coping with multiple agents which need to learn cooperative strategies.