Dynamic epsilone deep reinforcement learning method based on epsilone-greedy

Abstract

The invention discloses a method for deep reinforcement learning of dynamic epsilon on the basis of epsilone-greedy. The method comprises the following steps of: firstly, carrying out deep reinforcement learning on the dynamic epsilon; the method relates to the field of exploration and utilization dilemma, and comprises the following steps: preprocessing data of the dobby machine; according to the difference between the instant reward rt + 1 and the average reward R average, judging whether to update epsilon or not; a deep reinforcement learning framework is constructed by using a Dueling DQN algorithm; according to the TD-error of the behavior value function, judging whether to explore or utilize and update the value of epsilon at the moment; and importing the TD error in the experience pool into a Dueling DQN reinforcement learning framework, and training and updating the network. According to the invention, dilemma problem research of exploration and utilization is promoted; the exploration time and the utilization time are reasonably designed; according to the method, the final result is that dynamic updating of the epsillon is achieved, the dynamic updating problem of the epsillon can be achieved, and an effective theoretical basis is provided for development of the dynamic epsillon.

Description

technical field [0001] The invention belongs to the field of deep reinforcement learning and exploration and utilization dilemma, and relates to a method for deep reinforcement learning dynamic epsilon, in particular to a deep reinforcement learning dynamic epsilon method based on epsilon-greedy. Background technique [0002] Reinforcement learning is continuous learning in the process of interacting with the environment, and the quality of data obtained in the interaction largely determines the level of policies that the agent can learn. Currently, reinforcement learning (including deep reinforcement learning DRL and multi-agent reinforcement learning MARL) has excellent performance in the fields of games, robotics, etc., but nonetheless, in the case of reaching the same level, the sample size required for reinforcement learning ( interactions) is still far more than that of humans. This need for a large number of interaction samples seriously hinders the application of re...

AI Technical Summary

Problems solved by technology

[0003] In the classic multi-armed slot machine problem, the traditional epsilon-greedy has strong randomness. After continuous interaction, in order to get a higher return, the value of epsilon will become smaller and smaller, and the choice will be biased to obtain a large return. Action, if there are more rods, there will be a local optimal situation

Images