Utilizing unsupervised weightless neural network as autonomous states classifier in reinforcement learning algorithm

摘要

An implementation of reinforcement learning algorithm in an autonomous system requires knowledge expert to specify anticipated states, actions and rewards; and the algorithm will autonomously discover a near optimal behaviour for the system through trial-and-error interactions with its environment. The information on anticipated states are usually extracted from data streams and pre-programmed based on the knowledge expert interpretation of the data thus making the reinforcement learning algorithm rigid to only handles anticipated circumstances and the system will not be able to optimize. As an alternative, in this paper we explore the use of AUTOWiSARD, an unsupervised weightless neural network which will autonomously classify the states based on sensor information and then used by Q-learning, a reinforcement learning algorithm in order find near optimal behavior. The implementation will be demonstrated in an autonomous mobile robot simulation and the outcome will be presented and discussed.