Convergence estimation utilizing fractal dimensional analysis for reinforcement learning

Abstract

This paper proposes a novel convergence estimation method for reinforcement learning. In recent years, actual multi-robot systems utilizing reinforcement learning have been deployed in real-world situations. However, conventional learning methods require a substantial amount of time to reach convergence. Moreover, conventional learning processes are often inefficient because in most cases they are executed on a single robot only. In response to this problem, we propose a knowledge co-creation framework (KCF) for multi-robot systems, whose efficient implementation requires an autonomous convergence estimation method for reinforcement learning. Therefore, based on the assumption that learning curves exhibits fractality, we propose a convergence estimation method utilizing fractal dimensional analysis. Furthermore, we confirmed that the proposed method is capable of determining whether the learning would reach convergence by conducting a computer simulation.