Forbes reports a simulated aerial dogfight will be held next week and we can watch it live:
“The action will kick off Tuesday with AI vs. AI dogfights, featuring eight teams that developed algorithms to control a simulated F-16, leading to a round robin tournament that will select one to face off against a human pilot Thursday between 1:30 and 3:30 p.m. EDT. You can register to watch the action online. DARPA adds that a “multi-view format will afford viewers comprehensive perspectives of the dogfights in real-time and feature experts and guests from the Control Zone, akin to a TV sports commentary desk.
“With remarks from officials including USAF Colonel Daniel “Animal” Javorsek, head of the ACE program, recaps of previous rounds of the Trials, scores and live commentary, it’ll be just like Sunday Night Football — but on Thursday afternoon.Eric Tegler, “An Air Force Pilot Will Battle AI In A Virtual F-16 Dogfight Next Week. You Can Watch It Live” at Forbes
Note: Registration is now closed for non-U.S. citizens. The deadline for U.S> citizens is 4 PM EST August 17.
One goal of the contest, which will take place in a simulated environment, is to test AI. But DARPA (Defense Advanced Research Projects Agency) says the primary goal is to build “trust, particularly among U.S. fighter pilots, in artificial intelligence as the Pentagon seeks to develop unmanned systems that will fly and fight alongside them.”
One advantages of AI-controlled dogfights is a faster OODA (observe–orient–decide–act) loop during a conflict. As with gunfighters facing each other in a face-to-face quick draw contest in the old West, fast reactions favor success. Also, the effect of g-forces that black out pilots is no longer a concern in unmanned aircraft.
In engineering research parlance, the dogfight simulation is a proof of concept at the level of a toy problem. If ultimately deployed, unmanned AI fighter jets must ultimately be tested in diverse situations in the real world. Simulation is not enough.
Simulations are typically hard wired into computer code and can therefore be mastered using the same reinforcement learning used in training AI to beat the world champions at Go. Fixed simulations can be gamed in the same sense that Go has been gamed. The rules of performance, even if accompanied by randomness, can be looked at again and again. Repeatable access to a fixed set of rules allows reinforcement learning to develop strategies to use against the enemy—just as in the fixed game of Go.
If, however, the enemy gets a copy of the simulator and your system is trained only on the simulator, enemy AI can formulate counter strategies.
The real world is not like a simulation. In real scenarios, situations are invariably encountered that are not anticipated by simulations. Donald Knuth, prolific programmer and algorithm expert, said it well about his own computer code “Beware of bugs in the above code; I have only proved it correct, not tried it.”
Engineers have a similar saying: “In theory, theory and reality are the same. In reality, they’re not.”
As the complexity of a system increases linearly, its contingencies increase exponentially. What contingencies will accompany the AI jet fighter in the real world?
The Level 5 self-driving vehicle has encountered such unexpected contingency problems, delaying its development, which in turn has caused some AI companies to abandon the research. The development of self-driving cars that can navigate the country roads in the West Virginia mountains remains elusive. Time will tell.
Thursday’s AI dogfight in a simulated environment is a necessary step towards developing AI-controlled jet fighters. But, even if the concept proves successful, much work is needed before such fighters will be useful in combat.
Further reading: Calvin and Hobbes explain why AI will never rule the battlefield. The creativity needed for successful command is beyond the capability of AI.