A method and apparatus for reinforcement learning

Abstract

The invention discloses a method and apparatus for reinforcement learning. The method comprises the following steps: running a reinforcement learning model to obtain a first training data set; deriving a state transition function and a reward function of the reinforcement learning model by counting the first training data set; based on the state transition function and the reward function, solvingan optimal state estimation function through a relation reinforcement learning algorithm; modifying a reward value in the first training data set by using the optimal state estimation function; and training the reinforcement learning model by using the modified first training data set. According to the method, the final effect obtained by training the reinforcement learning model can be improved,and the training speed of the reinforcement learning model can be increased, so that the application of reinforcement learning in each actual scene is expanded.

Description

technical field [0001] The present invention relates to the field of machine learning, in particular to a method, device and storage medium for reinforcement learning. Background technique [0002] Reinforcement learning technology has achieved good results in many application fields. The deep reinforcement learning method based on neural network has achieved great success beyond the top human level in Atari video games, Go, Japanese chess and chess. But the training of deep reinforcement learning is very difficult. There are two main reasons. One is that the training signal of reinforcement learning is sparse throughout the training process. Specifically, only a small part of the training data has a reward function value that is not zero; the second is the role of the reward function Effects are usually delayed, i.e. actions that should be rewarded often get their corresponding reward signal many actions later. [0003] Traditional methods usually manually set the reward ...

AI Technical Summary

Problems solved by technology

The common feature of these methods is that not only a lot of labor is required, but also the newly set reward function is not completely equivalent to the original reward function in the environment due to human knowledge, which leads to the final training strategy and the actual optimal strategy. There is a deviation

Images