Robust reinforcement learning technique with bigeminal representation of continuous state space for multi-robot systems

Abstract

We have been developing a reinforcement learning technique called Bayesian-discrimination-function-based reinforcement learning (BRL) as an approach to autonomous specialization, which is a new concept in cooperative multirobot systems. BRL has a mechanism for autonomously segmenting the continuous state and action space. However, as in other machine learning approaches, overfitting is occasionally observed after successful learning. This paper proposes a technique to sophisticatedly utilize messy knowledge acquired using BRL. The proposed technique that has a doubly represented state space by parametric and nonparametric models is expected to show better learning performance and robustness against environmental changes. We investigate the proposed technique by conducting computer simulations of a cooperative transport task.