Knowledge Reusing Neural Learning System for Immediate Adaptation in Navigation Tasks

Abstract

A characteristic feature of conventional intelligent agents is the amount of trials that are required for them to learn. Since the tasks they encounter change depending on the environment, it is difficult for a learning system to compress the learning time using a priori knowledge. In the real world, however, agents confront a whole range of new tasks one by one, and have to solve them one by one without consuming learning time. A serious problem for the real-world agent is the amount of learning time needed. We suppose that one reason for a long learning time is the nonuse of prior knowledge. It is natural to expect that a kind of fast adaptation to tasks would be possible when we reuse knowledge that is acquired from similar past experiences. For this problem, we propose a neural-network based learning system that can immediately or quickly solve new tasks by reusing knowledge already acquired. We adopt a navigation task as an example and show the effectiveness of our method in variations on the task by comparing its performance with other methods.