Self-Regulation Mechanism: A Principle for Continual Autonomous Learning in Open-Ended Environments

Continual and autonomous learning are key features for a developmental agent in open-ended environments. This paper presents a mechanism of self-regulated learning to realize them. Considering the fact that learning progresses only when the learner is exposed to appropriate level of uncertainty, we propose that an agent's learning process be guided by the following two metacognitive strategies throughout its development: (a) Switch of behavioral strategies to regulate the level of expected uncertainty, and (b) Switch of learning strategies in accordance with the current subjective uncertainty. With this mechanism, we demonstrate efficient and stable online learning of a maze where only local perception is provided: the agent autonomously explores an environment of significant-scale, and self-develops an internal model that properly describes the hidden structure behind its experience.