杏吧原创

AN AI pilot has defeated a human 5-0 in virtual air-to-air combat in the first contest of its kind.

The AlphaDogfight Trials, staged by the US Defence Advanced Research Projects Agency (DARPA), were streamed online and proved so popular that the final had to be transferred to YouTube because of lack of capacity on the original platform.

While the US already operates many uncrewed drones, these are flown remotely by pilots, whereas simulated F-16 fighter aircraft in the AlphaDogfight Trials were controlled by artificial intelligence. In two previous events, AIs competed against each other. This time the organisers decided the technology was ready to take on a human.

Initially, eight teams from industry and academia went through three competition stages to find a machine champion, battling each other in various forms. This was won by AI specialists Heron Systems. Its AI then took on a human F-16 pilot known as Banger, who wasn鈥檛 named for security reasons. The AI beat Banger in all five encounters.

Every encounter started with the aircraft in random positions with neither having the advantage. The combats involved almost continuous turning, with both pilots pushing their aircraft to the limits to try to reach a firing position. This generally involved the aircraft chasing each other in tight circles.

The Heron AI was particularly aggressive in its manoeuvring, and, like the other AIs, showed no hesitation in flying at very low altitudes that human pilots tend to avoid. The human vs AI combats were notably shorter than many of the AI vs AI battles.

The Heron AI was, like most of the others, based on deep reinforcement learning, a technique in which algorithms learn through many attempts at a task. Unlike some others in the competition, it wasn鈥檛 combined with any pre-packaged responses, and the AI learned everything itself. Although some of the AIs updated their decisions up to 50 times a second, the Heron AI only updated 10 times a second to make its movements smoother.

This smoothness was particularly evident when Heron took on Lockheed Martin in the AI vs AI final. When the two aircraft were flying straight towards each other, the Heron AI could line up better to score more hits, destroying the other plane first.

The competition is part of DARPA鈥檚 Air Combat Evolution project, which envisages a future in which a pilot could have autonomous drones fly alongside them, or AI might manoeuvre an aircraft while the pilot selects targets. This would require a lot of confidence in the AI鈥檚 ability.

The competition is really to help pilots gain respect for the AIs, says Dan Javorsek, manager of the DARPA programme. 鈥淔ighter pilots have an almost institutional requirement to experience proof at the hands of one of their own before adopting a new tactic or technology.鈥�

He says the plan is to load algorithms into the flight control system of an aircraft. The type hasn鈥檛 yet been confirmed, but F-16s retired from front-line service have recently been converted so that they can be controlled remotely. The AI software converts them into autonomous dogfighters.

Dogfighting isn鈥檛 a big part of modern air combat, says Justin Bronk at defence think tank RUSI in London. Bronk says exchanges of fire are likely to start with radar-guided missiles well beyond visual range, with agile heat-seeking missiles used at closer range and guns as a last resort.

Bronk says that combat pilots require many other skills, for example, when they intercept an aircraft that isn鈥檛 responding to air traffic control and may be a threat. They need to be able to determine from context if it is a communications malfunction or a hijack situation. 鈥淔or example, is a pilot slumped over, or is there someone else in the cockpit?鈥� he says.

鈥淭he pilot needs understanding of the broader political context, whether the geopolitical background is normal or there is an elevated state of tension,鈥� says Bronk. 鈥淭his is needed to help classify what constitutes hostile behaviour, and how to respond appropriately.鈥�

Deploying AI-controlled fighter jets in action may be some way off, because the US Air Force is committed to maintaining human control over lethal systems to ensure responsibility and to prevent 鈥渦nwanted engagements鈥�.

The next stage of the DARPA programme will involve more complex actions with multiple aircraft on each side. This may give AI, which can easily track many objects simultaneously, even more of an edge over humans.