References & Citations
Computer Science > Artificial Intelligence
Title: Transfer Learning and Curriculum Learning in Sokoban
(Submitted on 25 May 2021 (v1), last revised 10 Sep 2021 (this version, v2))
Abstract: Transfer learning can speed up training in machine learning and is regularly used in classification tasks. It reuses prior knowledge from other tasks to pre-train networks for new tasks. In reinforcement learning, learning actions for a behavior policy that can be applied to new environments is still a challenge, especially for tasks that involve much planning. Sokoban is a challenging puzzle game. It has been used widely as a benchmark in planning-based reinforcement learning. In this paper, we show how prior knowledge improves learning in Sokoban tasks. We find that reusing feature representations learned previously can accelerate learning new, more complex, instances. In effect, we show how curriculum learning, from simple to complex tasks, works in Sokoban. Furthermore, feature representations learned in simpler instances are more general, and thus lead to positive transfers towards more complex tasks, but not vice versa. We have also studied which part of the knowledge is most important for transfer to succeed, and identify which layers should be used for pre-training.
Submission history
From: Zhao Yang [view email][v1] Tue, 25 May 2021 07:01:32 GMT (3857kb,D)
[v2] Fri, 10 Sep 2021 06:24:14 GMT (3764kb,D)
Link back to: arXiv, form interface, contact.