Hierarchization of Skill Acquisition without a Reward Function in Reinforcement Learning

Abstract

In recent years, many researches on pre-training have been done. For the pre-training in which a target task is unknown, some methods enable an agent to acquire skills without a reward function. These methods have a limitation that learning becomes unstable as the number of skills increases. In this research, we propose a model to acquire more skills while keeping diversity by hierarchizing skill acquisition. As a result, we achieved learning in a short time compared to existing methods for the same number of skills. Examination of multiple indicators showed that the diversity of skills is preserved. In addition, we confirmed that features of an agent state tend to be passed down through the hierarchical skills.