Multi-agent reinforcement learning method based on game reduction

Abstract

The invention discloses a novel multi-agent reinforcement learning method, and particularly relates to a multi-agent reinforcement learning method based on game reduction. According to the method, ina multi-agent task, the relationship among the agents is modeled and then grouped, and the agents in the same group are considered as a whole, so that information needing to be considered by the agents is simplified, and the learning efficiency of the agents is improved. Compared with an existing multi-agent reinforcement learning algorithm, the multi-agent reinforcement learning algorithm has theadvantages that the learning and decision-making speed of the agents can be increased to a great extent, and meanwhile, due to reduction of the learning complexity of the agents, the final learning effect of the agents is also improved. Experimental results show that the learning speed after game reduction is higher, and the final performance is better.

Description

technical field [0001] This patent discloses a new multi-agent reinforcement learning method, especially a multi-agent reinforcement learning method based on game reduction. Background technique [0002] Reinforcement learning is an effective technical means to solve sequential decision-making problems. It is a trial-and-error learning method. The agent learns the optimal behavior strategy by interacting with the environment, obtaining environmental feedback signals, and adaptively adjusting according to the feedback signals. , in order to maximize its long-term cumulative reward. In a multi-agent system, each agent not only interacts with the environment, but also cooperates or competes with other agents. The change of the environment state is no longer determined by a single agent, but by the behavior of all agents; at the same time, the reward signal obtained by each agent interacting with the environment also depends on the behavior of all agents. Since the agents are ...

AI Technical Summary

Problems solved by technology

Since the agents are all learning, the strategy of each agent will continue to change. For each individual agent, the state transition of the environment and the acquisition of reward signals lose their originality due to simultaneous learning by multiple agents. Some stability, which undoubtedly brings challenges to the application of reinforcement learning in multi-agent systems

Images