NEWS AI versus F-16 Pilot

[Jim123]

John Boyd would have gone into a negative g pull-over, flipped inverted, and snapped a picture.

Those were John Boyd's feet out the bottom of the one at the end. He was circling back for the kill shot.

 

[Swanee]

You're missing the point of the exercise, although everything you say is true.

In the old days of, for example, missiles, humans programmed the missiles to behave the way they do. A bunch of transfer functions and algorithms. So in a way, humans were still "flying" the missiles. You could work backwards from the maneuvers you saw to the lines of code and parameter settings that caused it.

With the AI, you just set it up to learn from scratch how to fly. It fails and crashes millions of times (in simulation) before it starts to succeed, then gets better, then gets more better. At no point in the training can you trace its actions in flight to weights in the neural network. It's pretty much BFM (not Basic Fighter Maneuvering either). It can learn things that humans haven't thought to try, and therefore wouldn't have thought to program in.

Obviously there were a ton of safeguards inserted into this thing, so its not a pure borg. But super cool nonetheless.

We did a project where we taught an AI-controlled glider to dynamically soar. It came up with a trajectory that no one had contemplated. Minds were blown.

1- AI glider? Interesting....

2- working on a project to develop autonomous recce. Big thing we're learning is that sensor control is hugely human centric. The way AI wants to do things is faster, more efficient, and much more accurate.

 

[SynixMan]

So it's doing a "gun face shot" at the merge? Keeping it UNCLASS, is that a thing?

If anything, I think this bodes well for a more complicated missile/expendable hit to kill vehicle that can be carried by our large supply of Gen 4.5 fighters. Launch HAL9000 into the fight and out-turn the hundreds of J-whatevers.

 

[Hair Warrior]

It can learn things that humans haven't thought to try, and therefore wouldn't have thought to program in.

Well, what you describe is machine learning (ML) - which is not quite AI, and DoD leaders often confuse the two and/or just lump them together to avoid mixing them up. Yes, ML makes AI better over time and more lethal, generally, unless it learns “bad” behaviors.

Watching that video, you could make the AI-piloted jet about 100x more lethal by installing rear-firing cannons in addition to the forward firing cannons. It would be absurd and damn near impossible to employ rear firing cannons effectively as a human, who would have to adjust the nose of the aircraft in an opposite direction to achieve a firing solution on a bogey at its 6 o’clock, but a computer could do it with relative ease.

 

[nittany03]

Watching that video, you could make the AI-piloted jet about 100x more lethal by installing rear-firing cannons in addition to the forward firing cannons. It would be absurd and damn near impossible to employ rear firing cannons effectively as a human, who would have to adjust the nose of the aircraft in an opposite direction to achieve a firing solution on a bogey at its 6 o’clock, but a computer could do it with relative ease.

"Donny, you're out of your element . . ."

 

[taxi1]

Well, what you describe is machine learning (ML) - which is not quite AI, and DoD leaders often confuse the two and/or just lump them together to avoid mixing them up.

Guilty as charged

There’s been a big effort to be able to come up with AI that can “explain itself” as to why it does what does. No luck so far. I don’t think they’ll ever be successful.

 

[Swanee]

Guilty as charged

There’s been a big effort to be able to come up with AI that can “explain itself” as to why it does what does. No luck so far. I don’t think they’ll ever be successful.

So many decisions are made, with so much information in. It would require a HUGE amount of data storage to be able to map out what Hal was thinking when it made one of a million decisions.

 

[nittany03]

So many decisions are made, with so much information in. It would require a HUGE amount of data storage to be able to map out what Hal was thinking when it made one of a million decisions.

The other thing I noticed was that (if I read it right) not only was this a sim, but they also gave the AI global SA as to what the fighter was doing. That's . . . not gonna happen in real life. You're going to have to rely on real-world sensor data that is incomplete and imperfect, if not being deliberately degraded. Wonder how much EA fuckery it can stand, and when, before the ML part starts being negatively trained. Remember when Microsoft tried to roll out Tay the bot, and she got trained into a flaming racist by Internet trolls?

 

[Jim123]

Remember when Microsoft tried to roll out Tay the bot, and she got trained into a flaming racist by Internet trolls?

Speaking of global SA...

?

Tay (chatbot) - Wikipedia

en.wikipedia.org

Trolling level achievement unlocked: Dungeon Master

 

[Hair Warrior]

Guilty as charged

There’s been a big effort to be able to come up with AI that can “explain itself” as to why it does what does. No luck so far. I don’t think they’ll ever be successful.

It's all good. I would have been just as guilty one year ago on explaining AI vs. ML.

The big breakthrough would be, suppose you have 30x fully autonomous air-to-air fighter jets in a combat theater. If one gets shot down 10 nm away from the rest of its fleet, as it goes down in flames do its "lessons learned" (on what not to do, or what the adversary TTPs were) get beamed back and incorporated into the neural network of ML that powers the algorithms of the 29 remaining jets? And if so, does their firmware get updated midair (near instantly), back on the CVN/tarmac, or only back at depot level maintenance? Now how about if it's shot down 100 nm away? 1,000 nm away? If you can solve that problem, then you are The Borg.

You get my point. If the "loser" algorithm gets obliterated by an AIM-9X before it can send its "lesson" back to the root software code that drives the rest of the AI jet algorithms, it's machine learning at less than 100%. Over time, it could be exploited because "winner bias" (Darwin?) would keep the winning software around longer and longer, but the loser software wouldn't quite be able to "learn from" why it lost.

But none of that matters if you can produce 20x AI jets for 1/4 the cost and 1/100 the time as it takes to make a single aviator and his/her F-35.

 

[Brett327]

high altitude, high speed, high endurance, high maneuverability, low RCS, low cost, small size, expendable, etc.

I take it you haven't met many engineers.

 

[Swanee]

And if so, does their firmware get updated midair (near instantly), back on the CVN/tarmac, or only back at depot level maintenance? Now how about if it's shot down 100 nm away? 1,000 nm away? If you can solve that problem, then you are The Borg.

ONR has projects out there that discuss just that.

How do we create the Cylons to be smart enough to kill anything and anyone, yet dumb enough not to question who is in charge.

 

[Griz882]

I take it you haven't met many engineers.

 

[Mos]

View attachment 27169

Without some sort of sub-rules for #1, it would seem to be difficult to make an AI do combat.

 

[Hair Warrior]

I take it you haven't met many engineers.

I have. Sounds like yours are just pessimists or have been too long in government.

How do we create the Cylons to be smart enough to kill anything and anyone, yet dumb enough not to question who is in charge.

Human-in-the-loop or human-on-the-loop