Active Bayesian feature weighting in reinforcement learning robot

A priori knowledge incorporation is known to be a bias for robot's exploration. This bias is intended to guide a robot to improve the quality of learning performance by selecting more significant sample in search space. However, main drawbacks of priori bias are that there is no guarantee that the final behavior is optimal and bias may be incorrect when the environment is changing. In this paper, we proposed additional guidance in a framework of active Bayesian network. Pre-defined features and expected utility function in our approach are used to determine the weighting factor of selecting action in Q-learning. We also use "evidence" from robot's experience which able to indicate that the current guidance knowledge (bias) is correct or not. This information is used to update parameters of Bayesian network by probabilistic adaptation algorithm. The posterior guidance knowledge can be taken into account based on this updating. Our approach is stated in general framework, which can be applied in any applications. A simple maze navigation problem is presented, using a Nomad200 mobile robot equipped with wireless video camera and frame grabber.