From pixels to policies: A bootstrapping agent

Abstract

An embodied agent senses the world at the pixel level through a large number of sense elements. In order to function intelligently, an agent needs high-level concepts, grounded in the pixel level. For human designers to program these concepts and their grounding explicitly is almost certainly intractable, so the agent must learn these foundational concepts autonomously. We describe an approach by which an autonomous learning agent can bootstrap its way from pixel-level interaction with the world, to individuating and tracking objects in the environment, to learning an effective policy for its behavior. We use methods drawn from computational scientific discovery to identify derived variables that support simplified models of the dynamics of the environment. These derived variables are abstracted to discrete qualitative variables, which serve as features for temporal difference learning. Our method bridges the gap between the continuous tracking of objects and the discrete state representation necessary for efficient and effective learning. We demonstrate and evaluate this approach with an agent experiencing a simple simulated world, through a sensory interface consisting of 60,000 time-varying binary variables in a 200 x 300 array, plus a three-valued motor signal and a real-valued reward signal.