A reinforcement learning system embedded agent with neural network-based multi-valued pattern memory structure

Abstract

This paper concerns about a way of intellectualization of robots (called "agent" here). Human learns incidents by own actions and reflects them on the subsequent actions as own experiences. These experiences are memorized in his/her brain and recollected and reused if necessary. This research incorporates such an intelligent information processing mechanism, and applies it to an autonomous agent that has three main functions that is, learning, memorization and associative recollection. In the proposed system, an actor-critic type reinforcement learning method is used for learning. For memorization, we introduce the chaotic auto-associative model that is proposed by Chartier, and that is also used like mutual associative memory system. Moreover, to deal with the increase of information, the memory part has an adaptive hierarchical layered structure of the memory module that consists of chaotic neural networks, especially for multi-valued pattern. Finally, the effectiveness of this proposed method is verified through the simulation applied to the maze-searching problem.