Learning in embodied action-perception loops through exploration

Although exploratory behaviors are ubiquitous in the animal kingdom, their computational underpinnings are still largely unknown. Behavioral Psychology has identified learning as a primary drive underlying many exploratory behaviors. Exploration is seen as a means for an animal to gather sensory data useful for reducing its ignorance about the environment. While related problems have been addressed in Data Mining and Reinforcement Learning, the computational modeling of learning-driven exploration by embodied agents is largely unrepresented. Here, we propose a computational theory for learning-driven exploration based on the concept of missing information that allows an agent to identify informative actions using Bayesian inference. We demonstrate that when embodiment constraints are high, agents must actively coordinate their actions to learn efficiently. Compared to earlier approaches, our exploration policy yields more efficient learning across a range of worlds with diverse structures. The improved learning in turn affords greater success in general tasks including navigation and reward gathering. We conclude by discussing how the proposed theory relates to previous informationtheoretic objectives of behavior, such as predictive information and the free energy principle, and how it might contribute to a general theory of exploratory behavior.