Multiple rewards fuzzy reinforcement learning algorithm in RoboCup environment

In order to achieve the competition tasks for multicooperating robots through learning, the paper discusses a kind of method that is designed for multi-agent systems (MAS), called the multi-reward fuzzy Q-learning algorithm (MRFQLA), which can be applied to the environment of the Robot World Cup Tournament (RoboCup). In MRFQLA., multiple reinforcement functions are established, based on the different characters of multi-agent systems. When the learning robot executes an action, these functions create multiple reinforcement signals that give the criteria of this action from different points of view. A Takagi-Sugeno (TS) model of a fuzzy inference system is built, which integrates these multiple rewards into one signal as the feedback of the learning robot. This method enhances the efficiency of learning because multiple rewards increase TD error and eliminates the conflict between the short-term target and the long-term one. Computer simulations in the RoboCup environment are shown and a discussion is given.