Multi-agent reinforcement learning method and system based on population training

Abstract

The invention relates to a multi-agent reinforcement learning method and system based on population training. The method comprises the steps of obtaining a first training set according to game videos;training the multi-layer full convolution LSTM network by using the first training set to obtain a first intelligent agent; utilizing the first intelligent agent to perform self-gaming, and obtaininga first population after a set time period; selecting a second intelligent agent, a first intelligent agent set and a second intelligent agent set from the first population; utilizing the first intelligent agent to fight with the selected three groups of intelligent agents at the same time, and storing and updating the first population until any one of the selected three groups of intelligent agents goes wrong, so as to obtain a second population; selecting a replacement agent from the second population to replace the battle agent to continue to fight with the first agent, storing and updating the second population, and obtaining a third population; and until the number of the agents in the third population reaches a preset value, outputting the first agent. According to the invention, the intelligent agent capable of simulating unmanned system combat command and control can be trained.

Description

technical field [0001] The invention relates to the field of agent training, in particular to a multi-agent reinforcement learning method and system based on population training. Background technique [0002] A milestone event in the field of intelligent agents in recent years is that Alphago, an intelligent agent based on reinforcement learning, defeated the top human Go players, thus making deep reinforcement learning the most likely path to artificial intelligence. The main technology used by Alphago is self- game. The dependence on the number of datasets can be reduced by self-play, and even surpass human experts, which is almost impossible for ordinary deep learning. [0003] In the era of intelligence in the future, a large number of robots that can replace human pilots to perform reconnaissance, strike, and confrontation tasks will be used in military warfare. This puts forward high requirements for the accuracy, timeliness, and effectiveness of the command and contr...

AI Technical Summary

Problems solved by technology

[0004] At present, there are relatively few studies on intelligent command and control training systems in robot confrontation scenarios, and there are relatively large limitations. The main problems include: heavily relying on training data , the amount of training data is small, and the winning rate of the agent after training with these small amounts of data is low

Images