Unsupervised Learning of Goal Spaces for Intrinsically Motivated Exploration


Nov 03, 2017 (modified: Nov 03, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Intrinsically motivated goal exploration algorithms enable machines to explore and discover a diversity of policies in large and complex environments. These exploration algorithms have been shown to allow real world robots to acquire skills such as tool use in high-dimensional continuous action and state spaces. However, they have so far assumed that self-generated goals are sampled in a specifically engineered space. In this work, we propose to use deep representation learning algorithms to learn a goal space, leveraging observations of world changes produced by another agent. We present experiments with a simulated robot arm interacting with an object, and we study how the performances of exploration algorithms on such learned representations relate with their performances on engineered representations. We also uncover a link between the exploration performances and the quality of the learned representation regarding the underlying state space.
  • TL;DR: We propose a novel Intrinsically Motivated Goal Exploration architecture with unsupervised learning of a space where goals can be sampled, and compare systematically various representation learning algorithms in this context.
  • Keywords: Intrinsic Motivation, Exploration, Multi-Goal Learning, Autonomous Learning, Developmental Learning