Intelligent decision-making method and device based on offline-online hybrid reinforcement learning

Abstract

The application relates to an intelligent decision-making method and device based on offline-online hybrid reinforcement learning. The method comprises the following steps: acquiring an offline data set generated by interaction between two parties in air game simulation confrontation; collecting offline training samples from the offline data set by using a hybrid sampling method, training a pre-constructed strategy network and a Q value network according to an offline reinforcement learning algorithm and the offline training samples, and taking the trained strategy network and Q value network as initial networks for online training; storing an online data set in an experience playback pool where the offline data set is located, collecting online training samples from the current experience playback pool by using a hybrid sampling method, training the initial networks according to an online reinforcement learning algorithm and the online training samples, and obtaining a trained intelligent decision-making model; and assisting a first unmanned aerial vehicle in intelligent decision-making in an air game simulation confrontation scene according to the intelligent decision-making model. The method can improve the reliability and accuracy of intelligent decision-making of the unmanned aerial vehicle.

Images