Towards effective learning in Monte Carlo Tree Search

蒙特卡罗树搜索中的有效学习

• 批准号: 556170-2020
• 负责人: Müller, Martin
• 金额: $ 2.91万
• 依托单位: University of Alberta
• 依托单位国家: 加拿大
• 项目类别: Alliance Grants
• 财政年份: 2020
• 资助国家: 加拿大
• 起止时间: 2020-01-01 至 2021-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=705936
• 关键词: Towards; effective; learning; Monte; Carlo

The combination of machine learning and heuristic search methods has achieved groundbreaking successes in recent years. Some of the best-known examples are the AlphaGo and Alpha Zero architectures recently developed by DeepMind. These approaches to computer problem-solving combine the search method of Monte Carlo Tree Search (MCTS) with the machine learning method of Deep Reinforcement Learning. We propose research which generalizes the Alpha Zero search and learning approach, and addresses some of its shortcomings: 1. generalizing learning in MCTS to domains without a perfect simulation model, and 2. performing an in-depth study of the positive interaction between learning and search in the context of MCTS, and of possible bottlenecks for learning.

近年来，机器学习和启发式搜索方法的结合取得了突破性的成功。一些最著名的例子是DeepMind最近开发的AlphaGo和Alpha Zero架构。这些计算机解决问题的方法将蒙特卡洛树搜索（MCTS）的搜索方法与深度强化学习的机器学习方法相结合。 我们提出的研究，推广阿尔法零搜索和学习方法，并解决其一些缺点：1。将MCTS中的学习推广到没有完美仿真模型的领域; 2.对MCTS背景下学习和搜索之间的积极互动以及可能的学习瓶颈进行深入研究。