AFRL-Funded AI Beats Fighter Pilots in Simulated Dogfights

London An Air Force Research Laboratory program initiated to help improve simulators may hold the key to one of combat aviation’s next big challenges—harnessing the potential of unmanned “swarms” and manned-unmanned teaming. The program, dubbed Alpha, began life as an attempt to improve the adversary algorithms in the Advanced Framework for Simulation, Integration and Modelling program, delivered by Boeing to AFRL in 2013. The fuzzy logic-based artificial intelligence—created by Psibernetix, a three-man Ohio-based company—has achieved remarkable successes in a recent series of demanding simulated battles alongside, and against, experienced human pilots. During a presentation Wednesday to Defence IQ’s International Fighter conference in London, Nick Ernest, CEO of Psibernetix, described how 12 fighter pilots had flown simulated air battles with and against Alpha-piloted platforms. In man-vs.-machine contests, the humans were given a representative simulated platform with good conventional fighter performance and AWACS support: Alpha operated with significant restrictions. “They turned us down to 1.9 radial Gs, which is pretty low for a fighter aircraft,” Ernest said. “And we were restricted to 0.19 max for linear Gs, which is about 75 percent of the acceleration of a Honda Odyssey minivan. Additionally, we had 200 knots less max speed, shorter-range missiles, and we didn’t have AWACS.” Despite the capability gap, Alpha scored numerous wins against the humans, in both four-vs.-two and two-vs.-two set-ups. Unlike many other artificial-intelligence projects, Alpha requires very little computing power—it runs on a $35 Raspberry Pi—and it not only learns for itself, its processes can be validated in seconds. Ernest cautioned that it is still early days for the project. “This is not reality,” he emphasised. “There’s a great deal of work to be done before it gets too exciting, but [the simulated battle results are] an interesting data point.”