Multi-agent reinforcement learning method based on inductive logic programming

Abstract

The invention relates to a multi-agent reinforcement learning method based on inductive logic programming. The method comprises the following steps: step 1, constructing a multi-agent system; 2, encoding the environment local observation information acquired by the intelligent agent and the received communication information into first-order predicate representation by utilizing micro-inductive logic programming; 3, performing inference decoding on the first-order predicate representation to obtain an action probability; 4, the intelligent agent selects an action according to the action probability and interacts with the environment; and step 5, evaluating actions selected by the intelligent agents by using the dominant function, and performing optimization updating according to an evaluation result until strategies of all the intelligent agents are converged. The multi-agent reinforcement learning method based on inductive logic programming has excellent performance in cooperation tasks of the agents, not only can learn a strategy close to the optimal strategy, but also has better interpretability compared with a traditional reinforcement learning method.

Description

technical field [0001] The invention belongs to the technical field of multi-agent reinforcement learning, in particular to a multi-agent reinforcement learning method based on inductive logic programming. Background technique [0002] Deep reinforcement learning has made breakthroughs in many fields such as human-computer games, video games, and real-time strategy games, showing amazing potential. Reinforcement learning focuses on how an agent should act in an environment to maximize long-term rewards. It is regarded as a promising approach and crucial to the development of general artificial intelligence. Most successful applications of reinforcement learning have been in single-agent domains, where modeling or predicting the behavior of other agents in an environment is largely considered unnecessary. However, there are many important applications in the real world that involve the interaction between multiple agents, and the emergent nature and complexity of the enviro...

AI Technical Summary

Problems solved by technology

The non-stationarity of this environment will make it difficult for the agent's strategy to converge

Second, each agent may have limited access to other agents' observed states

May cause agents to converge to suboptimal strategies due to inability to utilize information from other agents

Finally, the use of deep neural networks makes the policies learned by agents hard to interpret

[0004] Due to the problems of non-stationarity, partial observability, and interpretability in multi-agent reinforcement learning, there are many difficulties in the application of multi-agent reinforcement learning.

Images