Multi-agent group cooperation strategy automatic generation method

Abstract

The invention relates to the field of artificial intelligence, and discloses a multi-agent group cooperation strategy automatic generation method, which defines agents and strategy networks thereof according to a specific application environment, evaluates the networks and experiences, and realizes automatic generation of a multi-agent cooperation strategy. The adopted algorithm provides three innovations on the basis of the MADDPG algorithm: trace information, multi-agent cooperative team formation and birth and death training. The learning history of the intelligent agent in the environmentcan leave a trace amount of information in the environment, and the user can learn the experience of other people through the trace amount of information intelligent agent to avoid walking; the training efficiency can be improved through cooperative team formation of the multiple intelligent agents; finally, the agents with excellent learning ability in the environment are inherited to all information of themselves through filial generations to continue to be trained through birth and death training, the agents with poor learning ability in the environment return to the initial point to be trained again through death, and the generation and evaluation efficiency of the multi-agent cooperation strategy can be greatly improved.

Description

technical field

[0001] The invention relates to the field of artificial intelligence, in particular to a method for automatically generating multi-agent group cooperation strategies. Background technique

[0002] MADDPG is a multi-agent reinforcement learning framework based on the deep deterministic policy gradient algorithm, which can be used for the automatic generation of multi-agent cooperative strategies.

[0003] In a multi-agent system, each agent learns to improve its strategy by interacting with the environment to obtain a reward value (reward), so that the process of obtaining the optimal strategy in the environment is multi-agent reinforcement learning.

[0004] In single-agent reinforcement learning, the environment of the agent is stable, but in multi-agent reinforcement learning, the environment is complex and dynamic, which brings great difficulties to the learning process.

[0005] Dimension Explosion: In monolithic reinforcement learning, state-value funct...

Images