Pseudo-environment and battlefield situation strategy transfer technology based on adversarial discriminative migration method

Abstract

The invention provides a pseudo-environment and battlefield situation strategy transfer technology based on an adversarial discriminative migration method. An adversarial discriminative method used for regression migration is adopted, and the effectiveness of tactical strategies from the pseudo-environment to the real environment is evaluated. Through domain randomization, a real RGB camera is used for recording real combat scenes such as mountains, rivers and military bases, the strategies learned in simulation have high robustness, and the learned strategies are enough to be directly transferred to real combat scenarios; and through a simulation method, the cost of dependence on the real world during mimicry migration is greatly reduced.

Description

technical field [0001] The invention belongs to the field of artificial intelligence, and in particular relates to a mimetic environment and battlefield situation strategy transfer technology based on an adversarial discriminative transfer method. Background technique [0002] At present, in military units, war games are mostly used for turn-based tactical exercises. War game deduction cannot achieve real-time drills in actual combat exercises, and cannot truly simulate combat scenarios. Therefore, when conducting large-scale and long-term battles, war games can only focus on small-scale, short-term combat effects, and cannot achieve long-term tactical transfer. By building a combat mimetic environment, a real mimetic environment can be constructed through various sensors and intelligence data, and real-time tactical strategies can be generated to better respond to dynamic changes in the battlefield. [0003] In some environment-based training models, when the learning mod...

AI Technical Summary

Problems solved by technology

[0003] In some environment-based training models, when the learning model is applied to a real scene, if the scene is different from the scene configured for training, most current systems are very fragile

Some research methods use real environment data for model training, but the model collects real scene data in the real environment, which is very expensive

Images