Clicked:Curriculum Learning Connects Knowledge Distillation for Four-Player No-Limit Texas Hold'em Poker

Equilibrium solution is the core subject of imperfect information game. Although game solving paradigms vary from knowledge-guided tree search, through game reinforcement learning, to offline blueprint strategy pre-training and online adaptation, strategy transferability is the key to game solving,...

Full description

Saved in:
Bibliographic Details
Published inChinese Control and Decision Conference pp. 833 - 837
Main Authors Luo, Junren, Zhang, Wanpeng, Gu, Xueqiang, Wang, Zhangling, Chen, Jing
Format Conference Proceeding
LanguageEnglish
Published IEEE 25.05.2024
Subjects
curriculum learning
Estimation
Games
Hold'em poker
imperfect information game
knowledge distillation
Learning systems
Reinforcement learning
Online AccessGet full text
ISSN1948-9447
DOI10.1109/CCDC62350.2024.10587560

Cover

More Information
Summary:Equilibrium solution is the core subject of imperfect information game. Although game solving paradigms vary from knowledge-guided tree search, through game reinforcement learning, to offline blueprint strategy pre-training and online adaptation, strategy transferability is the key to game solving, that is, how to train robust response strategies for various types of opponents. In this paper, we propose a CLICKED method of curriculum learning connects knowledge distillation to solve multi-player imperfect information games. Different kinds of four-player curriculum are designed for curriculum knowledge distillation. Based on this framework, we designed a four-player Texas Hold 'em Poker agent that competed in the 2023 AAMAS Imperfect Information Game and won the third place. In addition, the experimental results prove the feasibility of the curriculum learning connects knowledge distillation method.
ISSN:1948-9447
DOI:10.1109/CCDC62350.2024.10587560