Group intelligence collaboration method for communication mechanism optimization

Abstract

The invention relates to a group intelligence collaboration method for communication mechanism optimization, and belongs to the technical field of group intelligence collaboration methods for communication mechanism optimization. The technical problem to be solved is to provide the improvement of the swarm intelligence collaboration method for communication mechanism optimization. According to the technical scheme for solving the technical problem, a centralized training and decentralized execution framework is adopted, a centralized evaluation network is adopted to measure the behavior value of the intelligent agent according to global state information, and the influence of the action of a single intelligent agent on the overall benefit is evaluated based on a time difference advantage strategy gradient; a shared storage communication mechanism is introduced into a strategy network part, local messages are generated for observation of a single agent, personalized global messages are generated for each agent through an attention mechanism and transmitted to strategy networks of the agents, and information exchange between the agents is completed; and the strategy network makes decisions according to agent local observation and global communication messages. The method is applied to the field of group collaboration.

Description

technical field [0001] The invention relates to a group intelligence collaboration method for communication mechanism optimization, which belongs to the field of multi-agent deep reinforcement learning. Background technique [0002] With the continuous advancement of science and technology, the combination of reinforcement learning (RL: Reinforcement Learning) and deep learning (DL: Deep Learning) has formed the field of deep reinforcement learning (DRL: Deep Reinforcement Learning). , automatic driving and logistics deployment and other scenarios, it shows a level beyond human beings. Due to the limited autonomous decision-making ability of a single agent, it often cannot have a great impact on the environment, and it is not suitable for scenarios that require multiple agents to cooperate to achieve a common goal. Therefore, researchers created the field of Multi-Agent Reinforcement Learning (MARL: Multi-Agent Reinforcement Learning) by applying DRL to multi-agent systems....

AI Technical Summary

Problems solved by technology

This type of method does not involve communication channels to complete the transmission of communication messages, resulting in the strategy network can only rely on its own observations to make decisions during the execution process, and cannot promote collaboration with other agents through communication

Images