Improved reinforcement learning with curriculum

作者：

Highlights：

• A RL neural-network/tree-search agent can be improved by using curriculum.

• End-game-first curriculum improves the quality of training buffer.

• Early-game experiences generated during early epochs are low quality.

摘要

•A RL neural-network/tree-search agent can be improved by using curriculum.•End-game-first curriculum improves the quality of training buffer.•Early-game experiences generated during early epochs are low quality.

论文关键词：Curriculum learning,Reinforcement learning,Monte Carlo tree search,General game playing

论文评审过程：Received 8 August 2019, Revised 19 February 2020, Accepted 2 May 2020, Available online 8 May 2020, Version of Record 29 May 2020.

论文官网地址：https://doi.org/10.1016/j.eswa.2020.113515