We present a methodology intended to optimize the flight path through a flight corridor occupied by enemy anti-aircraft guns. This is relevant for all kinds of aircraft, missiles, and drones moving through air space that is fully or partially controlled by such guns. To this end we use Q-learning - a type of reinforcement (machine) learning - which tries to find the optimal strategy to avoiding the anti-aircraft guns through repeated semi-random flight path trials. Q-learning can produce an optimal flight path through the enemy fire without modeling the anti-aircraft guns directly. An adversary response is still needed, but this can come from a black box simulation, user input, real data, or any other source. Here, we use an in-house tool for generating the anti-aircraft fire. This tool simulates a close-in weapons system (CIWS) guided by a fire control radar and Kalman flight path prediction filters. Q-learning can also be supplemented with neural networks - so-called deep Q-learning (DQN) - to handle even more complex problems. In this work, we present results for a subsonic flight corridor pass of one anti-aircraft gun position using classic Q-learning (no neural networks).
Optimizing Flight Paths Through Anti-Aircraft Gun Fire with Machine Learning
2021
6 pages
Report
Keine Angabe
Englisch
Aeronautics , Computers, Control & Information Theory , Aircraft fires , Aircraft guns , Anti-aircraft guns , Artificial intelligence software , Collision avoidance , Computers , Filters , Fire control radar , Flight paths , Guidance , Machine learning , Navigation , Neural networks , Radar , Simulations , Subsonic flight , Weapons