Top Gun was released 30 years ago and it looks as if the Maverick of tomorrow will be made of microchips. Developed by a University of Cincinnati (US) doctoral candidate, an Artificial Intelligence (AI) called ALPHA has consistently beaten other AIs and a retired United States Air Force Colonel in a high-fidelity, air-combat simulator using what's known as a genetic-fuzzy system that relies on off-the-shelf PC processors to do what was thought to be the reserve of supercomputers.
Unmanned Combat Aerial Vehicles (UCAVs) have made great strides in recent years, going from items of speculation to the decks of aircraft carriers. But however well they've done in taking off, landing, and carrying out assigned aerial missions, there's still been a big gap between what a human pilot can do and what a combat drone can hope to achieve. Until recently, experienced humans have found it easy to beat UCAVs in simulations after learning their tricks and weaknesses.
However, before opening the champagne and chalking up one for Team Homo Sapiens, remember that this isn't so much a victory as a problem. Modern fighter aircraft fly faster than the speed of sound at high altitudes in an environment filled with adversaries firing missiles that will soon go hypersonic. It's a blue sky battlefield where life and death is measured in microseconds – one where the limits of the human ability to assess and react are already being sorely tested.
So far, AI pilots haven't been able to do any better and often times fare worse than their human counterparts. This is because aerial combat requires processing so much information with so many variables affecting so many outcomes that even the best supercomputer can't keep up.
According to its makers at Psibernetix, Inc, the startup founded by UC College of Engineering and Applied Science 2015 doctoral graduate Nick Ernest, ALPHA may point the way to the artificial combat pilot of tomorrow. ALPHA is currently a research tool designed to test manned and unmanned teams in a simulated environment, but it is already consistently outperforming baseline computer programs used by the US Air Force Research Lab and human pilots.
In a recent test last October, Colonel Gene Lee, a highly experienced fighter pilot and trainer who has flown against AIs since the early 1980s, took on ALPHA in simulated dog fights. It was no surprise when ALPHA scored hits against Lee (AIs often do at first against humans), but the computer soon made clear that it could not only hit Lee, but that Lee couldn't score a kill in return as he was shot down time and again. Worse, (or better, depending on your perspective), ALPHA's success continued against Lee and other humans even when it was lumbered with digital handicaps, such as lower speeds and inferior weapons. And it did this in real time using US$500 computer hardware.
According to the University of Cincinnati, the key to ALPHA's success is what is called Genetic Fuzzy Tree (GFT), which is a subtype of fuzzy logic algorithms. Instead of dealing with all the logical outcomes of a situation in computations that could take years, ALPHA breaks things down into a set of sub-decisions and focuses only on relevant variables. To do this, it uses language-based programming instead of numerical, which makes it easy to improve the system and add expert knowledge.
New versions of the algorithm can be easily created at random or derived from older ones. These algorithms can compete against one another and the programmers can "breed" from the more successful ones. The result is an algorithm that has been compared to what Deep Blue vs. Kasparov was to chess.
Cincinnati sees ALPHA as the precursor to the artificial intelligence wingman of the future. By teaming artificial intelligence with US air capabilities, programs like ALPHA could lessen the likelihood of mistakes. With its ability to assess, plan, and respond to situations hundreds of times faster than a human pilot, it could make up for their slower reflexes, handle multiple targets, and manage weapons while allowing the flesh and blood commander to deal with higher-level decisions.
"ALPHA is already a deadly opponent to face in these simulated environments," says Ernest. "The goal is to continue developing ALPHA, to push and extend its capabilities, and perform additional testing against other trained pilots. Fidelity also needs to be increased, which will come in the form of even more realistic aerodynamic and sensor models. ALPHA is fully able to accommodate these additions, and we at Psibernetix look forward to continuing development."
The research was published in the Journal of Defense Management.
The video below shows the simulator in action.Source: