Improving the Robustness of Instance-Based Reinforcement Learning Robots by Metalearning

Abstract

Learning autonomous robots have been widely discussed in recent years. Reinforcement learning (RL) is a popular method in this domain. However, its performance is quite sensitive to the segmentation of state and action spaces. To overcome this problem, we developed the new technique Bayesian-discriminationfunction-based RL (BRL). BRL has proven to be more effective than other standard RL algorithms in dealing withmulti-robot system(MRS) problems. However, as in most learning systems, occasional overfitting problems occur in BRL. This paper introduces an extended BRL for improving the robustness of MRSs. Metalearning based on the information entropy of fired rules is adopted for adaptive modification of its learning parameters. Computer simulations are conducted to verify the effectiveness of our proposed method.