Log-learning: A preliminary study

摘要

The paper presents a preliminary study of the log-learning approach. Reinforcement learning literature typically discusses tabula rasa agents. While not having to depend on prior knowledge in general is useful, in many cases more complex tasks prove intractable when no knowledge is communicated to the agent in advance. The log-learning approach presents a way to encode knowledge of useful partial behaviours. The approach is verified and experimental results are given. The paper concludes by suggesting some further steps of inquiry.