On-line robot learning using the interval estimation algorithm

摘要

A lot of reinforcement learning algorithms are based on a full state space to learn from. In the RoboCup mid-size league this is impossible to do during the real games, due to the immense state space. This paper suggests a way to reduce the state space significantly by selecting among behaviors that are only triggered by few states. In fact to make the robot keeper learn very fast to select its best behavior with the purpose to defend the goal, we only used a single state in our experiments. For a behavior with a certain goal several implementations are made. From this behavior set the interval estimation algorithm chooses the behavior that has the highest probability to actually achieve the highest possible performance. This means fast learning, although the reduced state space also means that some solutions cannot be found.