| Popis: |
Multi-agent system is a distributed decision-making system composed of multi-agents interacting with environment, which is an important research direction of distributed artificial intelligence. Multi-agent system has broad application prospects in the complex and unknown real world, such as industry, agriculture, military, aerospace and other group robot systems, as well as traffic control, resource management, commercial finance and game AI. Multi-agent reinforcement learning (MARL) relies on the sequential decision-making ability of reinforcement learning in the unknown environment, and integrates many disciplines such as operations research, game theory and group psychology. It can give better play to multi-agent cooperative advantages and complete complex tasks with low cost and high efficiency. This article focuses on the analysis, comparison and prospect of the research results of deep reinforcement learning for multi-agent learning cooperative tasks. Firstly, it introduces the research background of MARL and the classification of learning tasks. Secondly, according to the key research content of MARL, MARL algorithm is divided into three cate-gories, including value decomposition, Actor-Critic, and experience replay. The differences of algorithms are compared from diverse perspectives such as environmental non-stationarity, credit assignment and convergence performance. Finally, some challenges faced by the future research of MARL are analyzed, and the application of MARL is prospected. |
