Transfer in reinforcement learning via shared features

摘要: We present a framework for transfer in reinforcement learning based on the idea that related tasks share some common features, and can be achieved via those shared features. The attempts to capture notion of are but distinct, provides insight into when usefully applied problem sequence it cannot. apply knowledge problem, show an agent learn portable shaping function from experience significantly improve performance later task, even given very brief training period. also skill transfer, agents skills across tasks, approaching perfectly learned problem-specific skills.