A fuzzy adaptive algorithm for expertness based cooperative learning, application to herding problem

Cooperative learning in multi-agent systems is generally expected to improve both quality and speed of learning. This is particularly true when agents are able to recognize expert agents amongst themselves and integrate their knowledge properly. Additionally, the process of learning can be improved when the reinforcement learning signals in each agent can balance between searching behavior of the unknown knowledge (exploration) and learning behavior of the obtained knowledge (exploitation). In this paper, a fuzzy dynamic cooperative learning method, based on weighted strategy sharing (WSS), is introduced which draws a balance between exploitation and exploration behaviors. In the weighed strategy sharing method, agents share their learned knowledge by a measure of their expertness. The strategy, when applied to the classic herding problem, shows further improvement in quality and speed of learning when parameters of the learning algorithm are dynamically determined by a fuzzy routine.