关键词: 博弈学习, 多智能体学习, 元博弈, 在线无悔学习

Abstract:

The new wave of artificial intelligence brought about by deep learning and reinforcement learning provides an "end-to-end" solution for agents from perception input to action decision-making output. Multi-agent learning is a frontier subject in the field of intelligent game confrontation, and it faces many problems and challenges such as adversarial environments, non-stationary opponents, incomplete information and uncertain actions. This paper starts from the perspective of game theory, firstly gives the organization of multi-agent learning system, gives an overview of multi-agent learning, and briefly introduces the classification of various multi-agent learning research methods. Secondly, based on the multi-agent learning framework in games, it introduces the basic multi-agent game and meta-game models, game solution concepts and game dynamics, as well as challenges such as diverse learning objectives, non-stationary environment (opponent), and equilibrium hard to compute and easy to transfer. Then comprehensively sort out the multi-agent game strategy learning methods, offline game strategy learning methods and online game strategy learning methods. Finally, some frontiers of multi-agent learning are discussed from three aspects of agent cognitive behavior modelling and collaboration, general game strategy learning methods, and distributed game strategy learning framework.

Key words: learning in games, multi-agent learning, meta-game, online no regret learning