Multi-agent confrontation method and system based on dynamic graph neural network

Abstract

The invention belongs to the field of reinforcement learning of a multi-agent system, particularly relates to a multi-agent confrontation method and system based on a dynamic graph neural network, and aims at solving the problems that an existing multi-agent model based on the graph neural network is low in training speed and low in efficiency, and much manual intervention is needed in graph construction. The method comprises the following steps: obtaining an observation vector of each agent, and carrying out linear transformation to obtain an observation feature vector; calculating a connection relationship between adjacent agents, and constructing a graph structure between the agents; carrying out embedded representation on a graph structure between the intelligent agents in combination with the observation feature vectors; performing network space-time parallel training on the action prediction result of the action network and the evaluation of the evaluation network by using the embedded representation; and performing action prediction and action evaluation in multi-agent confrontation through the trained network. According to the method, a more real graph relationship is established through pruning, space-time parallel training is realized by utilizing the full-connection neural network and position coding, the training efficiency is high, and the effect is good.

Description

technical field [0001] The invention belongs to the field of reinforcement learning of multi-agent systems, and in particular relates to a multi-agent confrontation method and system based on a dynamic graph neural network. Background technique [0002] Reinforcement learning has many successful applications in many fields, and multi-agent technology, as an important branch of reinforcement learning, has also been studied by many scholars. An important multi-agent research direction is multi-agent cooperative confrontation. Multi-agent models are mainly divided into two categories, one is confrontation and the other is cooperation. A typical confrontation model is AlphaZero, which requires two agents to play a game to find a better strategy. The cooperative relationship requires multiple agents to cooperate to complete tasks, such as tennis doubles, football tasks, etc. [0003] For multi-agent tasks, one of the most direct ideas is to directly equip each agent with a sin...

AI Technical Summary

Problems solved by technology

[0006] In order to solve the above-mentioned problems in the prior art, that is, the existing graph neural network-based multi-agent model training speed is slow, the efficiency is low, and the problem of requiring more manual intervention in graph construction, the present invention provides a dynamic graph-based A multi-agent confrontation method of a neural network, the multi-agent confrontation method includes:

Images