Fighter jet AI consistently beats “Top Gun” tactical experts

Callsign ALPHA, this high-fidelity air combat simulator was the victor in a simulated scenario against field experts.

0 June 28, 2016

With extensive aerial combat experience, retired U.S. Air Force Colonel Gene Lee has been flying in simulators against artificial intelligence (AI) opponents since the early 1980s. Lee explained that most experienced pilots are able to beat AI if they know what they are doing.


Retired United States Air Force Colonel Gene Lee, in a flight simulator, takes part in simulated air combat versus artificial intelligence technology developed by a team consisted of industry, US Air Force and University of Cincinnati representatives. Photo courtesy of Lisa Ventre, University of Cincinnati.

However, one particular AI simulator gave Lee a real challenge. Dubbed ALPHA, this high-fidelity air combat simulator was the victor in a simulated scenario.

“[ALPHA is] the most aggressive, responsive, dynamic and credible AI I’ve seen to date,” explained Lee.

Psibernetix Inc developed ALPHA and related tools. The company founded by UC College of Engineering and Applied Science 2015 doctoral graduate Nick Ernest, now president and CEO of the firm; as well as David Carroll, programming lead, Psibernetix, Inc.; with supporting technologies and research from Gene Lee; Kelly Cohen, UC aerospace professor; Tim Arnett, UC aerospace doctoral student; and Air Force Research Laboratory sponsors.

AI Simulation

The application features a genetic-fuzzy system specifically designed for use with Unmanned Combat Aerial Vehicles (UCAVs).

In early iterations, ALPHA easily beat other AI opponents. Lee repeatedly attempted to score a kill against more mature versions of ALPHA. However, the artificial intelligence combat simulator shot Lee out of the air every time during protracted engagements. ALPHA has bested Lee and other field experts.

“I was surprised at how aware and reactive it was,” said Lee. “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.”

Lee has trained with thousands of U.S. Air Force pilots, flown in several fighter aircraft and graduated from the U.S. Fighter Weapons School, yet when Lee flies against ALPHA in hours-long sessions that mimic real missions, “I go home feeling washed out. I’m tired, drained and mentally exhausted. This may be artificial intelligence, but it represents a real challenge.”

However successful ALPHA has been, Nick Ernest still sees room for improvement. “Fidelity also needs to be increased, which will come in the form of even more realistic aerodynamic and sensor models,” he explains.

In combat situations costly mistakes can be made in a matter of microseconds. ALPHA’s operations already occur significantly faster than do those of other language-based consumer product programming so the simulator will be important to lessen the likelihood of mistakes by operators. In fact, ALPHA can take in the entirety of sensor data, organize it, create a complete mapping of a combat scenario and make or change combat decisions for a flight of four fighter aircraft in less than a millisecond. Basically, the AI is so fast that it could consider and coordinate the best tactical plan and precise responses, within a dynamic environment, over 250 times faster than ALPHA’s human opponents could blink.

Programming Intelligence

ALPHA and its algorithms require no more than the computing power available in a low-budget PC in order to run in real time and quickly react and respond to uncertainty and random events or scenarios.

Ernest began working with UC engineering faculty member Cohen to resolve that computing-power challenge about three years ago while a doctoral student. The team tackled the problem using language-based control (vs. numeric based) and using what’s called a “Genetic Fuzzy Tree” (GFT) system, a subtype of what’s known as fuzzy logic algorithms.

“The easiest way I can describe the Genetic Fuzzy Tree system is that it’s more like how humans approach problems,” explained Ernest. He suggests that in a real game scenario, the player would not go through all of his/her opponent’s stats for the season, rather the person would make a judgement call, “my opponent is very good.” However,  the opponents historic capability wouldn’t be the only variable and these other factors would need to be taken into consideration.

That’s the basic concept involved in terms of the distributed computing power that’s the foundation of a Genetic Fuzzy Tree system wherein, otherwise, scenarios/decision making would require too high a number of rules if done by a single controller.

Added Ernest, “Only considering the relevant variables for each sub-decision is key for us to complete complex tasks as humans. So, it makes sense to have the AI do the same thing.” That’s the “tree” part of the term “Genetic Fuzzy Tree” system.

Generational Iterations

Ernest and his team developed AI algorithms that are language based, with if/then scenarios and rules able to encompass hundreds to thousands of variables. This language-based control or fuzzy logic, while much less about complex mathematics, can be verified and validated.

Linguistic control also allows for expert knowledge and advice to be imparted to the system. In this case, Lee worked with Psibernetix to provide tactical and maneuverability advice which was directly plugged in to ALPHA.

The ALPHA programming can also be improved from one generation to the next, from one version to the next.

“In a lot of ways, it’s no different than when air combat began in World War I. At first, there were a whole bunch of pilots,” explained UC’s Cohen. “Those who survived to the end of the war were the aces. Only in this case, we’re talking about code.”

ALPHA’s training has occurred on a $500 consumer-grade PC. This training process started with numerous and random versions of ALPHA. These automatically generated versions of ALPHA proved themselves against a manually tuned version of ALPHA. Successful strings of code are then “bred” with each other, favoring the stronger, or highest performance versions.

This is the “genetic” part of the “Genetic Fuzzy Tree” system.

“All of these aspects are combined, the tree cascade, the language-based programming and the generations,” said Cohen. “In terms of emulating human reasoning, I feel this is to unmanned aerial vehicles what the IBM/Deep Blue vs. Kasparov was to chess.”

Leave a Reply

Your email address will not be published. Required fields are marked *