Label prediction model training method, label prediction method and device

Abstract

The invention discloses a label prediction model training method, a label prediction method and a device, and is applied to the field of artificial intelligence. The method comprises the steps of: acquiring to-be-trained game data from a game video sample; acquiring image features according to the to-be-trained game data; acquiring in-game information features according to the to-be-trained game data; according to the to-be-trained game data, acquiring M overall-view labels, wherein M is an integer greater than or equal to 2; and according to the image features, the in-game information features and the M overall-view labels, training an overall-view label prediction model. According to the method, corresponding overall-view labels can be extracted from multiple categories for a same frameof training data, and the overall-view labels of multiple categories are combined to train the model, so that a single macroscopic intention is avoided, and the real intention of a player is describedmore perfectly. And the training data can be automatically labeled through a program, so that the labor cost of expert labeling is saved, and the labeling convenience is improved.

Description

technical field [0001] The present application relates to the field of artificial intelligence, and in particular to a method for training a label prediction model, a method and a device for label prediction. Background technique [0002] Artificial intelligence (AI) programs have already beaten professional players at board games with well-defined rules. In contrast, the operation of multiplayer online battle arena (MOBA) games is more complicated. For MOBA games, micro-operations mainly refer to the specific operations of game characters in the current scene, while the overall view mainly refers to the large-scale transfer and scheduling of game characters to form certain strategies and tactics. [0003] At present, a hierarchical macro strategy model based on MOBA games is proposed. The input of the model is attribute information (such as heroes, wild monsters, and soldier lines, etc.), map information (such as obstacles, etc.), global information (such as time, etc.) an...

AI Technical Summary

Problems solved by technology

[0004] However, the definition of the big picture label in the prior art is based on the location of the attacking resources. Therefore, in the micro-operation prediction process, only the macro strategy for the purpose of attack is considered, while in the actual operation of the player, the macro Intentions include more than attacks. The overall view of the label is one-sided, and it is difficult to fully describe the player's true intentions.

Images