A Discussion of reinforcement learning for similarity in state-action space

Abstract

To learn the action in a dynamic environment in robotics, reinforcement learning is a method to acquire the state-action space. However, if the complex environments and complex objectives is given, the number of states increases and it exponentially caouses many attempts of learning. It takes unterminated time for convergence of the learning. In this paper, we introduce geometrical similarity of conditions, discuss the efficiency of learning. Normally, in the state-action space, eligibility of the trace is well known to improve the efficiency of learning. In order to simplify the eligibility of traces, we apply the similaritiy under scaling and rotation in the state-action space. Currently, we experimented using only the similarity of symmetry, it shows the possibilities of our method.