Dynamic reconstruction of state space for behavior acquisition of reactive agent

摘要

When an autonomous agent learns goal-oriented behavior, whether it can utilize an appropriate state space or not is one of the most critical issues for its success. For this reason, several researchers have proposed methods of autonomous construction of state spaces. However, most of them have a disadvantage in requiring a tremendous amount of training data. In this paper, we propose a novel "re-construction" method, with which an agent can rebuild autonomously its manually defined initial state space based on the later experiences. This method reconstructs the state space by discretizing the continuous sensor space based on the entropy minimization of multiple behavior outcomes. As a result, autonomous agents can obtain appropriate state spaces much more efficiently and flexibly than by conventional construction methods.