Chinese AI team claims big win in battle to teach dogfights to drones
- System developed in Sichuan cuts ‘training’ time by learning 5,000 times faster than US counterpart
- Drones expected to play a bigger combat role will mean making computer chips work smarter
In domestic peer-reviewed journal Acta Aeronautica et Astronautica Sinica on Friday, the researchers said the higher learning speed could help the drone identify “cheating manoeuvres” by human pilots, reduce a computer chip’s workload and outperform opponents in complex, large-scale air combat.
“The algorithm in this paper can be extended to an air combat with multiple AI agents, which will be closer to the real situation in a battlefield,” the researchers at the China Aerodynamics Research and Development Centre in Mianyang, Sichuan province, said.
The researchers put the system to the test by simulating combat between a drone and a jet fighter.
In a similar dogfight competition in the United States in 2020, deep-learning AI systems were pitted against F-16 fighter jet pilots, with Maryland-based company Heron Systems declared the ultimate winner.
The Heron system defeated the pilots in all five dogfights, taking more than 4 billion rounds of “training” to achieve the result.
The researchers in Sichuan said their system took just 800,000 simulations to win most of its encounters.
In previous simulated combat, some experienced pilots used a deep dive as a last resort to send the drone crashing to the ground.
But the new AI foresaw the trap, pulled up at the last second while holding on to the opponent’s tail, according to lead author Huang Jiangtao and his colleagues.
Huang said the traditional approach to machine learning was inefficient because the systems simply repeated the training rounds “blindly”, with each new session based on random data generated by previous exercises.
Because most of the data was not useful, it could take a computer a very long time to make a leap in performance.
Huang said that their new AI system was selective, choosing only the best data for the next round.
Though the deep reinforcement learning model they used was more or less similar to those used by Heron and others, the selective algorithm made a big difference in the AI’s learning curve.
After less than 80,000 rounds of training, the machine was already flying steadily like a pro, according to Huang.
In one simulated battle using J-10 fighter jets, a pilot managed to gain an advantage in the first 30 seconds, then the chase reversed suddenly and for more than 12 minutes he could not shake off the machine.
Drones are expected to play a bigger part in defence.
Chinese military researchers have said that the country’s existing defences are vulnerable to relatively cheap US drones using next-generation stealth technology, which could cripple the Chinese military’s centralised command systems.
In response, the People’s Liberation Army has started decentralising some of its combat forces and ensuring flexibility against a future US attack with the help of similar AI-powered drones, according to air force researcher Major General Fei Aiguo.
But the AI system needed to improve its ability and learn faster from the rapidly changing environment, Fei and his colleagues said in a paper published in the journal Command Information System and Technology in October.
Most military drones have been designed for surveillance, early warning, communications or to attack ground targets. These systems cannot handle fast-paced, sophisticated actions such as dogfights because an enormous amount of calculation must be done quickly, according to some military researchers.
Another challenge is that most military computer chips were built with slower, conservative technology for harsh conditions such as extreme heat, pressure and electromagnetic jamming.
A main focus of the Chinese military AI programme is to develop new algorithms that can get high performance out of a slow computer.
Some recent breakthroughs in this effort allowed the Chinese military to apply the technology on some of its most advanced weapons that were previously believed to be too challenging for AI such as hypersonic platforms.
The AI would, for instance, allow a missile flying at five times the speed of sound in the air to hit a target with unprecedented accuracy, or a hypersonic plane to land in an airport with lower risk of crashing, according to researchers involved in these projects.