[1Q2-J-2] Machine learning: reinforcement learning and its advances

[1Q2-J-2-02] Human Sub-goal Transfer in Hierarchical Reinforcement Learning

〇Takato Okudo^1,2, Seiji Yamada^2,1 (1. SOKENDAI, 2. National Institute of Informatics)

Keywords:Hierarchical Reinforcement Learning, Human subgoal transfer, Interactive Machine Learning

Hierarchical reinforcement learning, especially which learn policy with option discovery simultaneously, needs a lot of iterations. This paper investigates how human sub-goal transfer affect to learning speed and performance. we proposes the way to transfer human sub-goals in hierarchical reinforcement learning. To acquire human sub-goal knowledge, we use the problem in interactive machine learning. Supervised learning transforms human sub-goals into initial parameters before learning on hierarchical reinforcement learning. Two experiments, participant experiment and evaluation experiment, are conducted. The participant experiment is to acquire sub-goals of participants. The human sub-goal transfer is evaluated on learning speed and performance after learning in evaluation experiment. The future work is to conduct two experiments and analyze the results.