Reinforcement learning in the robocup-soccer keepaway

摘要

Many researchers purpose reinforcement learning (RL) as a form of machine learning for robot learning. However, there are several issues that need to be considered when applying (RL) techniques to robot tasks. There are many different (RL) algorithms available such as Q-learning or Sarsa. These algorithms may produce different results. In complex domains with large states and action spaces is necessary to apply generalization techniques such as function approximation. Last, a right balance between exploration and exploitation is required. In this paper we review these issues in order to improve the learning process in the keepaway domain. We present some new combinations in the choice of the RL algorithm, the generalization method and the exploration-exploitation strategy.