Multi-agent game AI design method based on attention mechanism and reinforcement learning

Abstract

The invention discloses a multi-agent game AI (artificial intelligence) design method based on an attention mechanism and reinforcement learning, which comprises the following steps: firstly, performing entity-based local information division and grouping on agent original observation information; secondly, obtaining a local strategy at the current moment based on each piece of grouping information representation, obtaining an attention weight corresponding to local information according to the action-observation historical information, and aggregating all the local strategies according to the attention weight to obtain a local information strategy; and finally, aggregating the local information strategy and the original observation information as a reference strategy of strategy input and output to obtain a final strategy of the intelligent agent. Compared with the prior art, the method has the advantages that the problem of agent observation information redundancy in a multi-agent system is solved, and the learning efficiency of the agents can be effectively improved.

Description

technical field [0001] The invention relates to the field of multi-agent deep reinforcement learning, in particular to a design method for a multi-agent game AI in a combat environment. Background technique [0002] Multi-agent reinforcement learning is a very important research field in the field of artificial intelligence. It has been used to solve complex multi-agent problems such as cooperation of a large number of robot systems and automatic driving, and has achieved good results. [0003] In the field of multi-agent reinforcement learning, most algorithm frameworks adopt the principle of centralized training & distributed execution. Each agent independently uses its own observation information to make decisions during the execution process, thereby alleviating the impact on the policy network caused by the exponential increase in the joint observation space caused by the number of large-scale agents; in the training process, the joint actions of all agents are used- T...

AI Technical Summary

Problems solved by technology

The action space is divided based on action semantics. This division method is equivalent to artificially constraining the strategy of each local observation, which may lead to insufficient strategy fusion. At the same time, the difference in the size of the action subspace divided in the article leads to the final The complexity of the policy fusion process increases

[0008] At present, only a small number of studies have mentioned the impact of irrelevant information in the agent's observation information on the current decision-making, how to learn to judge the influence of each part of the information in the current observation information on the final decision-making, and strengthen the relevant information as much as possible and weaken it. Irrelevant information remains an open issue

Images