Game strategy optimization method and system and storage medium

Abstract

The invention provides a game strategy optimization method, a game strategy optimization system and a storage medium. The game strategy optimization method comprises a step of establishing a strategygradient algorithm based on maximum entropy and a step of solving a multi-agent optimal reaction strategy. The method has the beneficial effects that a centralized training and decentralized executionmode is adopted, the accuracy of an action valuation network is improved, Meanwhile, global baseline rewards are introduced to more accurately measure the action income of an intelligent agent, so that the problem of credit allocation in a human game is solved. Meanwhile, a maximum entropy method is introduced for strategy evaluation, and exploration and utilization in the strategy optimization process are balanced.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a game strategy optimization method, system and storage medium based on multi-agent reinforcement learning and virtual self-play. Background technique [0002] Many decision-making problems in real-world scenarios can be modeled as strategy solving problems in incomplete information games, but the current machine game algorithms need to abstract the state space of the problem, do not perform well in high-dimensional action spaces, and usually only It is suitable for two-player games, but most of the games in practical problems are multi-player games. [0003] Neural Fictitious Self-Play (NFSP) is a game strategy solving method that has attracted a lot of attention in the field of machine games. It learns through self-play, and uses deep reinforcement learning and supervised learning to realize machine games. Computation of the best response strategy and update of...

AI Technical Summary

Problems solved by technology

In NFSP (Neural Network Virtual Self-Play), each generated sample corresponds to a fixed opponent's strategy. If the strategy influence brought by other agents cannot be perceived, then the learned optimal response strategy and average strategy are inaccurate. , will cause the Markov property of the MDP (Markov decision process) to fail

In addition, in the multi-agent game, there are many problems such as dimension disaster, credit allocation, global exploration, etc.

Images