An AI program developed with funding from the US military has managed to consistently defeat human pilots in a training simulator. ALPHA, which was designed as a research tool, was called "the most aggressive, responsive, dynamic, and credible AI I’ve seen to date" by United States Air Force Colonel Gene Lee, a retired aerial combat instructor who tested the program's capabilities.
This was not dogfighting — but something much more tactical
Lee battled ALPHA in a flight simulator which pitched teams of aircraft against one another in various scenarios, including defending and attacking specific territory. These sorts of aerial combat situations are not dogfights (which are fought at close range) but long-range, tactical battles, where enemy aircraft are too far away to be seen with the naked eye and fight with missiles. In Lee's match-ups with ALPHA he was not able to score a single kill against the program, and said it represented a significant improvement on previous AIs.
"Until now, an AI opponent simply could not keep up with anything like the real pressure and pace of combat-like scenarios," said Lee in a press statement. "I was surprised at how aware and reactive it was. It seemed to be aware of my intentions and reacting instantly to my changes in flight and my missile deployment. It knew how to defeat the shot I was taking. It moved instantly between defensive and offensive actions as needed."
Both Lee and ALPHA had access to the same information during the battles, including data from various sensors, mixed in with random noise and potential instrument failures to make the simulation more realistic. Nick Ernest, CEO of Psibernetix, the company that developed ALPHA, told The Verge that while the simulator can still be improved in terms of making it more life-like (such as including more realistic modeling of aerodynamics), the AI will be more than capable of taking these changes on board.
In a paper published in the Journal of Defense Management, Ernest and his fellow researchers describe how they developed ALPHA using a machine learning technique known as genetic fuzzy tree methodology or GFT. This is a combination of two AI development methods: one, genetic algorithms, in which competing systems are pitted against each other in a manner analogous to evolutionary competition; and two, fuzzy logic, which mixes logical systems of calculation with probabilistic factors — hence the fuzziness.
ALPHA can run on nothing more than a "low-budget PC"
As well as its ability to learn, ALPHA is also notable for its lightweight performance. The program's creators say it requires no more computing power than a "low-budget PC" to run real-time simulations. They say considering this alongside its tactical ability and fast reaction times could make it an ideal controller for uncrewed aircraft. "Combining these strengths in a mixed manned and unmanned fighter squadron could prove to be an extremely effective fighting force," write the researchers. "ALPHA-controlled aircraft would happily volunteer to take risky tactics and have the manned craft perform safer support roles."
And, of course, an AI pilot will never get tired or need a break. Lee said that flying against ALPHA for hours at a time in the simulator was enough to prove how useful this particular skill is. "I go home feeling washed out," he said. "I’m tired, drained and mentally exhausted. This may be artificial intelligence, but it represents a real challenge.”