Partitioning Sensorimotor Space by Predictability Principle in Intrinsic Motivation Systems

Abstract

Inspired by infant development, intrinsic motivation (IM) guides the robot with intelligent exploration strategies, enabling efficient and effective learning in high-dimensional search spaces. A particular method in IM, namely Intelligent Adaptive Curiosity (IAC), adaptively partitions agents sensorimotor space $(\mathrm{S}\mathbb{M})$ into regions of exploration, and guides the agent to select the regions that are in the moderate level of difficulty, and learns separate experts for different regions. Therefore, the means of partitioning the $\mathbb{SM}$ and the mechanisms behind region generation is of utmost importance. In this study, we propose a method for partitioning the space that allows maximizing the performances of the experts that will be responsible for learning skills. In brief, for each potential partitioning, the error of the experts are calculated and the partitioning that would generate the minimal error in the future is selected. Our method is evaluated in a setting with a simulated robot that learns predicting the next state given the current state and the action taken in an environment composed of regions with different properties. We verified the proposed method, SM is partitioned into more semantically meaningful regions adapting environment dynamics, the exploration of the robot in these regions can better exploit IM mechanisms and the system learn more efficiently and effectively i.e. with higher performance in a shorter time, compared to a baseline method.