Learning to select relevant perspective in a dynamic environment

When an agent observes its environment, there are two important characteristics of the perceived information. One is the relevance of information and the other is redundancy. The irrelevant and redundant features which commonly exists within an environment, commonly leads to agent state explosion and associated high computational cost within the learning process. This paper presents an efficient method concerning both the relevance of information and the correlation in order to improve the learning of reinforcement learning agent. We introduce a new concurrent online learning method to calculate the match count C(s) and relevance degree I(s) to quantify the redundancy and correlation of features with respect to a desired learning task. Our analysis shows that the correlation relationship of the features can be extracted and projected to concurrent biased learning threads. By comparing the commonalities of these learning threads, we can evaluate the relevance degree of a feature that contributes to a particular learning task. We explain the method using random walk examples and then demonstrate the method on the chase object domain. Our validation results show that, using the concurrent learning method, we can efficiently detect redundancy and irrelevant features from the environment on sequential tasks, and significantly improve the efficiency of learning. After relevant features are extracted, the agent can remarkably accelerate its succeeding learning speed.