An Enactive Self-Model for Sparse Representations and Improved Performance

Justin Brody

Abstract: We present a machine learning model for an agent in a dynamic environment to learn a model of its body and actions. We test our model in the context of playing Atari games (Breakout and Asteroids), specifically by modifying Google DeepMind's well-known hierarchical Q-Network. We demonstrate that, compared to the control, our model learns a qualitatively sparser set of features, attains proficient game-play more quickly, and usually scores more points.