Training method and device for reinforcement learning model in battle game

Abstract

The invention provides a training method and device for a reinforcement learning model in a battle game, belongs to the technical field of computers, and relates to artificial intelligence and computer vision technologies. The method comprises the steps that a target battle model and a similar opponent model of the target battle model are acquired, the similar opponent model is a historical battlemodel with the grade score difference between the similar opponent model and the target battle model smaller than a score threshold value, and the grade score is used for evaluating the battle ability of the model; based on the battle state characteristics of the two battle parties, the prediction operation of the target battle model and the prediction operation of the similar opponent model aredetermined respectively; the target battle model and the similar opponent model are used for controlling the two battle parties to execute prediction operation so as to conduct battle; an operation value of the target battle model in battle is determined; and the target battle model is trained based on the battle state characteristics, the prediction operation and the operation value.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to a method and device for training reinforcement learning models in battle games. Background technique [0002] MOBA (Multiplayer Online Battle Arena, multiplayer online tactical arena game), also known as ARTS (Action Real-Time Strategy, action real-time strategy game). The gameplay of this type of game generally requires the purchase of equipment during battles. Players are usually divided into two or more hostile camps and compete with each other on a scattered game map. Each player controls the selected virtual character through the interface to fight against each other. fight. In the game, the virtual characters of the two camps can be controlled by the player to fight, or the players and AI (Artificial Intelligence, artificial intelligence) can control the virtual characters of different camps to fight. [0003] Due to the unbalanced strength and difficulty of h...

AI Technical Summary

Problems solved by technology

[0003] Due to the unbalanced strength and difficulty of heroes in the multiplayer game design, this imbalance can easily lead to an unbalanced lineup. In the traditional reinforcement learning training process, a certain probability is used to self-play or randomly select opponents from the historical opponent pool. This training method will cause weak lineups to be suppressed by strong lineups when training AI, resulting in a significant decline in AI's exploration capabilities, poor model training efficiency, and AI capabilities that cannot meet expectations

Images