Learning to solve complex tasks by reinforcement: a new algorithm

Abstract

In this paper, a new approach for learning to solve complex problems by reinforcement is proposed. In order to solve complex tasks the system is guided by a teacher who previously proposes intermediate general tasks to learn. The learnt behaviors to solve these tasks are added to the system's set of actions increasing its skills until it is able to easily solve the desired complex task. This approach uses a new reinforcement learning mechanism, robust to ambiguous information and able to learn general behaviors. These mechanisms are studied, described and finally tested with a set of experiments in a complex environment.