AI Beats Top F-16 Pilot In DARPA Dogfight Simulation

I watched a video of an F18 driver's analysis of the exercise. Among his observations:

Pilot had zero time in this particular sim.

National Guard pilot - not knocking the individual or his mode of service, simply was observed that this was "average joe" fighter pilot.

Video host F18 pilot suggested that the noise, Gforces, visual, and other physiological cues that are core to fighter pilot training, are absent in this type of Sim. That pilot would have been significantly impaired in ability to judge what maneuvers his opponent was engaging in, whether opponent energy increasing or decreasing, etc.

That military never trains dogfighting skills and techniques in sims, only in flying aircraft, due to the importance of these inputs to decision making processes.

A hit was only "was the gun lined up and in range," the AI did not have to click the trigger... effectively infinite ammo stream from AI guns

Sim had relatively poor fidelity even for a sim.

AI was "Omnipotent" basing decisions on receiving real time inputs of point in space, velocity directions, energy state changes, speeds, etc, 100% in real time and 100% of the exercise duration, as though it had a direct data flow from the opposing aircraft. An advantage that just is not realistic even if an AI was flying a real dogfight against an opposing real aircraft driven by a human. Whereas, Pilot had nothing other than his eyes to judge what his opponent was doing.

 

I'm probably incorrect or remembered incorrectly about some of these points. If you're groaning at my descriptions, I'm just a guy who watched a video a week ago!

 

On avg, most fighter pilots are avg (relatively speaking)

All the same information could be given to a pilot on a head-up display via the same sensors that the AI has, but then the pilot would have to look at the data, assimilate it, and then act on it. Looking at the data on a head-up would require the pilot to take their eyes off the target, something you can't do in a close in dogfight (note: they do have helmet mounted displays, although not sure how prevalent they are on fighter aircraft, if at all currently). The time to look/consider/act versus how fast a computer AI could perform the same functions could be the difference in the outcome.

 

I think you are missing the point...the pilot not having a data stream of information available as was available for the AI platform is unimportant. What's important is the simulator deprived the pilot of the real world cues that are roughly equivalent to that stream. When we fight we can judge rates of change in a variety of parameters with our eyes, muscles, and brains. Changes in the enemy's position relative to our head position and other cues in the cockpit, for example, do not require us to have constantly avaialable digital data on the enemy's flight position and parameters as was available to the AI platform. Depriving the pilot of the human equivalent of that data stream, if the reports are correct, was a setup for failure.

 

Obviously you have real world insight/experience here, but I don't think I am missing the point entirely. I agree that being in a sim would cause the human to be at a deficit, even IF they had a lot of time in a sim under these types of training scenarios. However, your analogy wrt to the data stream seems to me to be an analog vs digital comparison. Digital will win.

 

Now I don't get it then. I thought you were advocating streaming data to a pilot through the HUD or helmet mounted displays in order to make the match up more equitable. I was simply pointing out that the pilot only needed what would be normally available to him in a real F-16 to make the match equitable...no need for any new tech. If the report is correct, the pilot was severely hand-cuffed and criteria for a win for the AI was overly sympathetic. More bogus grant money results...

 

yes, I was responding to the statement (opinion by an observer???) that the AI had an advantage due to available data.

Don't disagree with that at all, except whether it would truly be equitable. The only true test is to fly it. Sim is not real. But I think a human using visual and physical cues will ultimately be slower reacting than an AI using sensor data.

Would that surprise you? Not really any different than many of the early SDI tests. Any new tech requires an increasingly greater level of sophisticated tests to vet and refine it.

 

Agreed. Particularly on the last point. If you've ever read some of my other points on manned versus unmanned fighters, you'll remember I'm a firm believer that we fighter pilots are soon to be shown the door. There are a million other ways to break other people's stuff with many that are way quicker and far more accurate.

 

Don't say that. My BIL, who was a Top Gun instructor, has a son (Jr) who just entered fleet transition training(?) at Lemoore, whose wife just gave birth to a son (III). I'm sure Sr and Jr already have visions of III following in their foot steps (both Academy grads and Naval aviators).

 

Image Unavailable, Please Login

Not desirable

 

Isn't this the same as "trusting" online poker sites to play fair?