Knowledge Transfer from a Human Perspective

Transfer in reinforcement learning (TiRL) is a challenging research problem. Agents are still preprogrammed for specific tasks or can only transfer knowledge under limited circumstances. To quicken the learning process, people, who are extremely adaptable, may be able to provide feedback to guide the agent. This requires an interpretable medium for transfer that both the human and agent can understand. In this work, we propose a transfer medium based on object mappings between tasks. We conduct human subject experiments to test whether people are able to effectively use these mappings in the form of advice to play video games. Preliminary results show that good mappings improve people’s transfer performance on some games, but can hurt people when they misunderstand. The potential interpretability benefits of using an object-based representation for TiRL can guide the development of more interpretable transfer learning algorithms for agents.