Abstract: XCS classifier system has been shown to solve machine-learning problems in a competitive way.However,in multi-robot problems,XCS is restricted to solve very small problems modeled by a Markov decision process.In this paper a new learning technique XCSG that combines XCS and gradient descent methods was proposed to solve multi-robot machine-learning problems.XCSG builds low-dimensional approximation of the function,and gradient descent techniques use on-line knowledge to establish a stable approximation of functions,so that the Q-form has been maintained at a low-dimensional stable state.Approximate of the function not only requires smaller storage space,but also allows the robot online knowledge is summarized on the generalization.Simulation results show that XCSG algorithm solves the multi-robot reinforcement learning in a large space,slow learning,learning uncertainty and other issues.

Key words: Reinforcement learning,Multi-robot,Accuracy-based learning classifier system(XCS),Accuracy-based learning classifier system with gradient descent method(XCSG)