In this paper a revised reinforcement learning method is presented for stability control problems with real-value inputs and outputs. The revised eXtended Classifier System for Real-input and Real-output (XCSRR) controller is designed, which is capable of working at fully real-value environment such as stability control of robots. XCSRR is a novel approach to enhance the performance of classifier systems for more practical problems than systems with merely binary behaviour. As a case study, we use XCSRR to control the stability of a biped robot, which is subjected to unknown external forces that would disturb the robot equilibrium. The external forces and the dynamics of the upper body of the biped robot are modelled in MATLAB software to train the XCSRR controller. Theoretical and experimental results of the learning behaviour and the performance of stability control on the robot demonstrate the strength and efficiency of the proposed new approach.
- eXtended Classifier Systems (XCS)
- real-value problem
- reinforcement learning
ASJC Scopus subject areas
- Computer Science(all)