Data processing method and device for game

Abstract

The invention provides a data processing method and device for a game. The method comprises the steps that multiple game features of a game image of at least a current frame of the game are acquired;at least multiple game features are input into a second processing network and output from the second processing network for evaluating an evaluation value of the action output by a first processing network, wherein when multiple game features of the game image of one frame is input into the first processing network, the first processing network outputs the to-be-executed action in the game imageof the frame; a reward value obtained by a game character in the process of performing the action in the game is acquired; based on the reward value, the second processing network is trained; based onthe evaluation value output by the second processing network, the first processing network is trained.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, more specifically, to a data processing method and device for games. Background technique [0002] With the development of multimedia technology and the popularization of wireless networks, people's entertainment activities have become more and more abundant, such as playing stand-alone or networked games through handheld terminal devices. There are various types of games. For example, a currently popular game is a First Person Shooting Game (First Person Shooting Game, FPS), in which the shooting game is performed from a subjective perspective of a player. [0003] In addition, with the development of artificial intelligence technology, the concept of game AI (Artificial Intelligence, artificial intelligence) is proposed. By training game AI, it can perform various actions as game characters in the game, so as to provide players with more powerful and realistic virtual enemies, a...

AI Technical Summary

Problems solved by technology

However, the problem is that, first of all, the strategy gradient method needs to complete a game before it can correctly calculate the value function used to evaluate the action of the network output, so it cannot be trained online in real time

[0006] Moreover, the processing network training based on the DQN algorithm and the PG algorithm all use images as input, so the state space is huge, resulting in a huge amount of game AI training

Additionally, this makes the processing network weights heavy, making it very CPU and GPU demanding during training

Images