Acceleration of Reinforcement Learning by a Mobile Robot Using Generalized Inhibition Rules

Abstract

One very fundamental problem in behavioral learning by an agent is that it takes quite a long time to acquire optimal behavior. In order to solve this problem, in this paper, we propose an approach to make learning processes more efficient by the use of generalized knowledge. In this approach, the agent repeats learning processes for different tasks and extracts behavioral rules that are commonly harmful to task execution by the use of statistical method. After sufficient experience is accumulated, the generalized rules are extracted from the experience and are applied to subsequent learning processes, and, consequently, the learning processes are accelerated by inhibiting commonly harmful behaviors. In order to achieve generality of rule expression, the description of the rules is based on egocentric information, namely, raw data of observations and actions experienced by the agent. In order to avoid a perceptual aliasing problem, the rule expression includes information on sequential experience and a mechanism is introduced to control the balance of utility and generality of the rules. The proposedmethod is examined in navigation tasks by amobile robot in grid environments as an example of application. The results show that the proposed method accelerates learning processes.