Air countermeasure maneuver control method based on random neural network

Abstract

The invention relates to the technical field of machine learning, in particular to an air countermeasure maneuver control method based on a random neural network. The method comprises the following steps: S1, executing a fighter maneuver decision based on a constructed vacuum confrontation model, interacting with a simulation environment, and collecting training sample data; S2, fusing environment external rewards and decision internal rewards to reconstruct a reward mechanism; and S3, training the air confrontation maneuver control method according to the collected training sample data and the reconstructed reward mechanism. Based on the air countermeasure maneuver control method, a fighter can achieve a good maneuvering control effect in an air countermeasure scene, and the migration generalization ability of an algorithm is improved.

Description

technical field [0001] The present application relates to the technical field of machine learning, in particular to an air counter maneuver control method based on a stochastic neural network. Background technique [0002] In the field of air confrontation, smart maneuver control method is the key to defeating the enemy. In the traditional research on fighter maneuver control, most of them are based on the existing maneuver control theory to strengthen the simulation and debugging of the fighter model. This type of maneuver depends to a large extent on the fidelity of the fighter model and human's prior knowledge of maneuver theory, and its flexibility and controllability are poor. With the continuous deepening of artificial intelligence research, fighter maneuver control is also developing towards a more intelligent direction. Among them, reinforcement learning can enable the agent to learn the strategy of obtaining the maximum reward during the interaction with the enviro...

AI Technical Summary

Problems solved by technology

Traditional exploration methods often perform poorly on this task, exhibiting high sample complexity

There are usually two ways to solve the problem of sparse rewards: one is to layer actions, and combine low-dimensional actions into high-dimensional meta-actions to reduce the search space, but requires more domain knowledge and careful design; the other is to use the intrinsic Rewards guide agent exploration without domain knowledge, but poor transferability

Images