Reinforcement learning method and related device

Abstract

The invention discloses a reinforcement learning method and a related device, a regularization item used by the reinforcement learning method is a regularization item based on a sample, the regularization item based on the sample comprises additional rewards for executing different actions by a target intelligent agent, the availability degree of exploration behaviors is defined, mutual influence between executed actions is modeled, a probability density function does not need to be calculated when the regularization term based on the sample is used for reinforcement learning, the reinforcement learning efficiency can be improved, and particularly the reinforcement learning efficiency for complex strategies is improved. In addition, the sample-based regularization term is used for reinforcement learning, so that geometrical information of an action space can be utilized, and the method can be compatible with a wider strategy structure.

Description

technical field [0001] The present application relates to the technical field of machine learning, and more specifically, relates to a reinforcement learning method and related devices. Background technique [0002] Reinforcement Learning (RL) has achieved great success in fields such as video games and robot control. The goal of reinforcement learning is to find an optimal policy by maximizing the cumulative reward, which usually leads to a deterministic policy. In practical application, the effect of adopting random strategy is better than that of adopting deterministic strategy. For example, for new tasks, stochastic policies tend to allow for better exploration; moreover, using stochastic policies tends to be more robust than deterministic ones when unexpected situations occur. [0003] To learn stochastic policies, existing learning methods suffer from multiple deficiencies, among which the main issue includes the low learning efficiency for complex policies. Conten...

AI Technical Summary

Problems solved by technology

[0003] In order to learn stochastic policies, existing learning methods suffer from multiple deficiencies, among which the main problem includes the problem of low learning efficiency for complex policies

Images