Artificial intelligence fought with a real pilot: the US Air Force conducted an imitation of air combat

In the USA, a virtual tournament was held between the F-16 Viper fighter with an artificial intelligence “pilot” and the best fighter pilot of the American Air Force.

According to The Drive author Joseph Trevithick, the simulated aerial combat is the culmination of recent efforts by the Defense Advanced Research Projects Agency (DARPA), launched a year ago to explore the automation and use of artificial intelligence in air combat. - air ".

The first place in the tests was taken by Heron Systems, which employs only 30 employees. She managed to bypass Aurora Flight Sciences, EpiSys Science, Georgia Tech Research Institute, Lockheed Martin, Perspecta Labs, PhysicsAI and SoarTech. The tournament lasted three days - from 18 to 20 August 2020 inclusive.

On the first day, 8 teams played against 5 different types of opponents, modeled by the Applied Physics Laboratory (APL) at Johns Hopkins University. Opponents included those who operated as fighters, heavy bombers, cruise missiles and large drones.

On August 19, the number of participants decreased and they began to act against each other. Four finalists have been identified: Aurora Flight Sciences, Heron Systems, Lockheed Martin and PhysicsAI. In the semi-finals, Lockheed Martin defeated Physics AI, and Heron Systems defeated Aurora Flight Sciences. And finally, the main thing happened: the Heron Systems team met in battle with a real pilot - an instructor on the F-16.

The tournament was the final in a series of events that began back in November 2019. True, Joseph Trevithick does not yet understand how the outcome of the competition can affect the implementation of the large-scale Air Combat Evolution (ACE) program. But DARPA believes the event will energize attendees and expand the "AI" developer base for the ACE program.

While such aerial combat is less likely in the future, the need to be able to handle situations like this is not going anywhere. We continue to use it in more complex scenarios such as suppressing enemy air defenses or in offensive air attacks,

- said in an interview with reporters, US Air Force Colonel Daniel Javorsek.

According to Trevithick, the first phase of the ACE program in the American military department is planned to be completed in the coming 2021. It will include flight tests of experimental systems controlled by artificial intelligence. DARPA expects these trials to test different types of propeller-driven and jet-powered unmanned aircraft capabilities. Then the systems will be tested on larger air vehicles.

The software generated from the program could enter the United States Air Force around 2024. As a result of the implementation of the program, aircraft will become more autonomous, and artificial intelligence will make it possible to make quick decisions in a dynamic air combat environment.

Joseph Trevithick concludes that digital aerial combat is now a testament to the ever-growing interest of the US military in artificial intelligence. It is especially important to assess the results of a virtual battle, where artificial intelligence fought with a real pilot.

Most notably, Heron Systems' algorithms were able to compete with a real Air Force fighter pilot. But experimental tests and competitions are one thing, and the real practice of air combat is another. But as a test of the capabilities of artificial intelligence with the further improvement of work on the automation of aircraft control, such tests are indeed of great interest.

Comments

[Mavrikiy]

Artificial intelligence fought with a real pilot: the US Air Force conducted an imitation of air combat by a pilot: the US Air Force conducted an imitation of an air battle

Reminds: "Artificial intelligence fought with a real chess player.

artificial intelligence will make it possible to make quick decisions in a dynamic air combat environment.

In battle, situations will arise that are not provided for by the program and ..... and it's too late to enter.

[Temples]

Who won?

[Mavrikiy]

Quote: Temples

Who won?

"Experience, son of difficult mistakes and genius, friend of paradoxes"

[Shurik70]

If AI had won, they would have said.
I remember the history of the late 80s. There one (American) boasted that his "scientific" group was studying the possibility of telepathic communication in practice. I didn’t say whether a telepathic connection was made even to the next room behind the wall. To a question from the audience "and how are you doing?" - replied that they conducted an experiment - one subscriber climbed in a balloon to the height of Everest, another in a bathyscaphe was under water at a depth of a kilometer ...

[hydrox]

Absolutely idiotic article :: it is not clear what and it is not known to anyone - neither the source code, nor the results ...

[Rusticolus]

As who？ As always, shkololo came with cheats and piled on everyone.

[Charik]

what kind of cheats, endless missiles or immortality

[Masha]

Quote: Temples

Who won?

but it doesn't matter ... it's important that ...

the event will allow to energize the participants and expand the base of "artificial intelligence" developers

and money will flow like a river ...

[APASUS]

Quote: Temples

Who won?

[voyaka uh]

Like who? Watch the video. Very interesting.
The final is at the end.
Heron defeated instructor "Banger" 5-0 dry. Five clean wins
in dog fights, closest fights.
The instructor had only one chance to hit - and then he missed.
Heron wrote out very beautiful thoughtful-calculated loops and
clearly went to the point of the shot.
I did not expect anything else. Computer reaction speed - thousands of times
higher than human reaction.

[PROXOR]

NUANCE! This is when the AI ​​continuously sees the enemy's plane, as it was in a virtual environment. The very first maneuver on real fighters and AI will lose the enemy. And while he gropes for him with cameras, he will not leave a wet spot from him. 5: 0 here is spite.

[Blackmokona]

360 degree view already implemented on the F-35

[PROXOR]

Well, well, well. Tell me how he is watching the enemy there. You can follow it visually or by means of radiation.

[Blackmokona]

That's it

The F-35 DAS OS will be capable of a 360-degree view with high refresh rate, high resolution and sensitivity, all projected onto the pilot's helmet display. The system consists of several optical sensors located at different points in the fuselage. The computer combines their images into one seamless picture of their surroundings. DAS detects and tracks targets in a completely passive mode (no radar or laser illumination), and does not require pilot intervention. As soon as the enemy appears on the battlefield, the DAS will instantly capture the target (ground, air, missiles, including air defense and air-to-air) and will continuously monitor it, excluding the possibility of leaving the blind spots. In this case, the pilot can shoot into the rear hemisphere and perform any maneuvers.
Source: http://nevskii-bastion.ru/an-aaq-37-usa/ VTS "NEVSKY BASTION" AVKarpenko

[akarfoxhound]

Advertising (especially minke whales) and reality outside the aircraft are most often mismatched things. Probably because of the "steepness" of this system, we have not yet seen the close air combat of the F-35 with the F-16, right?

[Blackmokona]

How bae have long seen

[hydrox]

Of course!
That's why American super-dupers are embarrassed to appear in the sky even on the edges of the screens of observation stationary radars, not like in the low-power radars of our front-line aviation ...
This artificial intelligence tells the pilots where and what ambit lies in wait for them ... especially if they become aware that the Su-35 are on patrol - they are immediately blown off the screens like a feather (excessive consumption of fuel, diapers and wild consumption of a resource on a long afterburner)

[_Ugene_]

yes, in fact, who would doubt it, creativity is not required here, a strict calculation of probabilities and speed of reaction, which means that in this area a person has no chance against AI

[Swordserg]

Neural network. With a score of 5: 0

[rocket757]

In the USA, a virtual tournament was held between the F-16 Viper fighter with an artificial intelligence “pilot” and the best fighter pilot of the American Air Force.

I remembered the cartoon "Miracles on the bends", how there Baloo lost the competition to the "damn shelezyaka"!
All pilots were immediately retired !!! but not for long, by the way, because the shelezaks in the end turned out to be no ice!
There will be good helpers for the crew, but ... not right tomorrow. And so, progress cannot be stopped !!!

[dauria]

In battle, situations will arise that are not provided for by the program and ..... and it's too late to enter.

What's the difference between a program in a car and a program in a brain? No principled. Man is the same AI. Something is in the ROM (genes), the rest is received after birth, then "renewal", experience. Where did you get the idea that the computer cannot learn? Without a doubt, in a dozen years nothing will shine on a person on board an aircraft. "The violinist is not needed, Uncle Vova. He only consumes excess fuel"
Although to land the F-15 with half a wing, as the Jewish pilot did - the question Well, it's not the gods who burn the pots. The computer will want to live - it will not be able to.

[hydrox]

It will not be able to: the parameters of such a composition still need to be entered into memory so that the AI ​​knows which direction to run with such indicators of flight parameters.
It's like in chess: there are trillions of combinations and you can brute force to find a winning one, but the transience of modern aerial close combat may not give a computer such an opportunity ...

[Ka-52]

all artificial intelligence is based on the same algorithms that are used in our country both in semi-natural simulations of air battles and in programming the ACS of combat fighters. For example, PfM in automatic mode in 4+ fighters has not caused any particular problems for a long time.

[voyaka uh]

You are wrong. AI is not based on the same principles as aircraft software
4th and even 5th generation.
This is a fundamentally different programming.

[Soho]

You are wrong. AI is not based on the same principles as aircraft software
4th and even 5th generation.
This is a fundamentally different programming.

Don't write nonsense. First, tell us how the AI ​​is programmed. What is taken as a basis for describing a mathematical model of flight, assessing the spatial position of an aircraft and other processes The mathematical models are the same. Are you raving about Cameron's Terminator films? It would be better to read what is taught at the Zhukovsky Air Force Academy. There, a whole course is devoted to this. The level of knowledge is a sofa, but you climb to put grades

[Shahno]

Your truth. Only I doubt that you, for example, will raise your logic adaboost to C ... And so yes, the principles were laid down by comrades Church and Turing ..

[Shahno]

I don't know what is now at the Zhukovsky Academy ...
But even we have an understanding that the algorithms of our developments have gone far from expert systems. That is, the models are not just different from each other, but a fundamentally different approach for a different project.

[akarfoxhound]

The Zhukovsky Academy promised to live for a long time under Tolyasik Taburetkin, moving to Monino, and then "to the south." In Voronezh, except for the name (like Gagarinskaya), there was nothing left, the optimizers were ruined, the scientific base was zero. And I doubt it will recover.

[Shahno]

Yes, it's sad. Did not know.

[eklmn]

“Don't write gibberish. First, tell us how the AI ​​is programmed. "
What is incomprehensible to yourself always causes such a reaction. Remember the reaction of the Inquisition to scientific progress. And how AI is programmed, before asking the forum - rummage yourself on Google, while it was closed for you.
From the article below:
"It took Lockheed Martin several servers running round-the-clock dogfights by trial and error to create their ultimate AI. a piece of software capable of running on a single video card.
As Defense One was told by Ben Bell, Senior Machine Learning Engineer at Heron Systems, The winning team's AI went through over 4 billion simulations, or about 12 years of flight. "

https://futurism.com/the-byte/ai-fighter-jet-obliterates-human

[Soho]

Unlike such gorlopanov like you, I have at least seen and am familiar with the method of calculating the components. And you like me only from Wikipedia can pull squeezes

[Shahno]

The people are just not quite .. as if they understand, they live at the level of 20 years ago ... And who is really interested can read the information on "q learning", "markov decision process", "gradient boosting". I apologize for the English, but that's how it worked out. Maybe I'll open a good translation into Russian, then I'll throw it off .. There are many interesting things in Kaggel, there are enough Russian guys.

[eklmn]

Thanks for the links, looked at Kaggel .. !!! Wow, but not for me anymore - not young already ...
I was doing dinosaur programming
Cobol, I understand a little the logistics (and only!) Of creating AI - it's damn hard! And as an engineer from Heron System writes, going through 4 billion simulations is a hell of a lot! Tesla has problems with AI in unmanned driving, but Tesla, if mistaken, really kills, she still has to work with AI and work ...
But in the article in the VO it is not written, but in the original there is that today American pilots are hostile to AI in their plane and / or in their wingman.
On a subconscious level, they find it difficult to believe that the AI's response / choice will be better than the pilot himself. These trainings are good for future military pilots.

[Soho]

that today American pilots are hostile to AI on their plane and / or in their wingman.

For a person who knows about combat aviation only from the pages of the yellow press such as VO, I report that aircraft control elements have existed in automatic mode since ancient times. ABSU is able to perform from civilian trajectory control to military PfP. In any military aircraft, AI elements are already at the SPG LEVEL !!!!
Another mattress idolater

[eklmn]

"For a person who knows about combat aviation only from the pages of the yellow press like VO, I report ..."
Yes, I've never been a pilot, but I'm interested. Here is a list of my “yellow press” (all in English):
https://news.usni.org/
https://nationalinterest.org/
https://www.defensenews.com/
https://futurism.com/the-byte
https://spacenews.com/
https://theaviationist.com/
https://www.defenseworld.net/
https://www.popularmechanics.com/military/

and many others. I recommend reading before bed - you will find many new and interesting things that broaden your horizons ...

"Another mattress idolater"
So after all I am a “mattress”, and I do not hide it .. !!

[Ka-52]

Yes, I've never been a pilot, but I'm interested

Well, unlike you, I speak from the original source.
Once again, for those who have difficulty comprehending what they have read: it is impossible to teach an AI to control an aircraft (even its simulation), if you do not have mathematical models of aircraft with their complex unstable separation-vortex flow structure. The same applies to weapons. If there is no algorithm for calculating the spatial position from the data obtained from onboard tracking devices, then the software complex will not be able to calculate the% of the likely destruction of the aircraft at the current coordinates. It will not be possible to calculate that point in space, upon arrival at which a miss with the highest probability will be provided. Therefore, it is necessary, without pigs' squeals of joy, to realistically assess what has been achieved by American engineers and programmers. It is necessary to understand that the elements of the elements of the program calculation have been and have been for a long time. Including on Russian technology. AI simply operates with them like bricks, trying to create a battle picture that has a maximum% of probable victory. But AI can never win a real war. Because operates only with the basis of tactical situations, which is spelled out in it. And a real battle does not always go according to the canons.

So after all I am a “mattress”, and I do not hide it .. !!

well, that doesn't mean you have to be a fool. After all, there are smart US citizens

[Shahno]

No approach to software is fundamentally different. The difference is primitively speaking between Java and Python .... Supervised reinforcement learning is going well. Pure NP comlexity level problem. With chess, it's a little different ...

[Pavel57]

In a real battle, there will be no time to learn artificial intelligence, but at the stage of development, as far as possible, everything can be foreseen.
Although human lost to AI in chess,

[Grazdanin]

Quote: Pavel57

Although in chess man lost to AI

The main question is which person. If the best losers, what about the middle ones? Quasi-AI can be scaled quickly and massively, unlike humans.

[keeper03]

AI is removed from the battle by special means and the aircraft turns to stone! But a trained pilot just can't be defeated !!!

[PROXOR]

In chess, the number of moves is limited by the pieces on the board and the AI ​​has no problem calculating them. Plus, all pieces move the same way. The dog dump is not chess. Fighters are different in their technical characteristics. Plus, now the AI ​​will not be able to track the rapidly changing environment in the real world. There are no such cameras and sensors. Well, the quality of the pilots of mattresses was shown in 1993, the demonstration battles of the SU-27 of the Lipetsk training center and the aces on the Fu-15 needles. As we know, they completely merged ours.

[iouris]

Just like Kasparov. Did you hear the mate?

[voyaka uh]

The instructor pilot was shocked, it seemed to me.
I'm sure that at the end - after the 3-4th defeat - he swore in earnest ...

[boss]

A swarm of identical planes with AI, over time, even if not today, are guaranteed to guide the pilot to the target.
If you need it at all.
After quality, quantity rules.
Which we don't have (

[g1washntwn]

For AI, an "identical plane" is no longer suitable, for it to be effective it needs to withstand well beyond 9G, but at a price it is desirable much lower. These are consumables for cases when he himself attacked "a little with su".

[ved_med12]

The first place in the tests was taken by Heron Systems, which employs only 30 employees.

And as a result, it will be absorbed by one of: Aurora Flight Sciences, EpiSys Science, Georgia Tech Research Institute, Lockheed Martin, Perspecta Labs, PhysicsAI and SoarTech.

[bar]

Artificial intelligence fought with a real pilot

Super. But after reading the article, I still did not understand who won whom.
However, this is a common thing for the author's articles.

[g1washntwn]

They said - friendship won. The tournament was held in order to feel fresh ideas and approaches. If they find something interesting for DARPA and private traders, they won't even tell them. Ideas for the little things will be copied under the modified recipe. That would be something new and breakthrough, patents, contracts and grants for development would immediately fall down, and so ... thank you all, everyone is free, go get energized until the next meetings.

[keeper03]

The collective mind has won !!!

[voyaka uh]

Defeated Human by Heron Program 5: 0

[Zaurbek]

Apparently, "smart wingmen" will puzzle in the future not only ground air defense, but also fighters.

[Grazdanin]

Yet these are different projects. In the project, loyal followers will retain the main computational tasks of the fighter. The task is to make relatively cheap drones for aircraft escort, not ultimate vehicles.

[Zaurbek]

Maneuverability should be better

[hydrox]

So, if the AI ​​hears the screeching of the guidance locator along the Birch, then it will be able to work out only one command: "Tick!" - and what will be the results of such an air battle?

[Zaurbek]

If Talalikhin writes the program, then it's not a fact

[hydrox]

And who attacked Talalikhin?
Chopping the tail with a screw is his personal decision and it remains to be seen whether the AI ​​will be able to make the same decision ...
So after all, the implementation of such a maneuver, and the results of such a maneuver are very ambiguous, there is no need to talk about the expediency of such actions at all ...

[Zaurbek]

I ran out of cartridges .... I chopped with a screw.

[hydrox]

You try to program the AI ​​for this position - it will be funny!

[Darkspy]

In 5 out of 5 battles, the AI ​​won the victory.https: //naked-science.ru/article/hi-tech/video-ii-pobedil-letchika-v-virtualnyh-vozdushnyh-bitvah

[Pavel57]

Quote: keeper03

The collective mind has won !!!

Collective but artificial.

[tralflot1832]

I didn't understand anything. The pilot and the intellect flew on airplanes or only on a komputore. Toda oh.

[Avior]

flew on chairs

Children's toy

[tralflot1832]

And of course, here someone wrote at the "open door" day. On the Kuznetsov simulator, a girl schoolgirl made a successful landing and this is from the same opera.

[Ryaruav]

to defeat a trained modern man at this historical stage of human development, even covid 19 is not possible while the world nature is fighting us in our arms

[Avior]

would write simply, who won and with what result.

[viktor_ui]

Artificial intelligence will make it possible to work at extreme overloads of the aircraft, safe for itself and lethal for the pilot. Waiting for the positronic brain from Isaac Asimov We are on the crest of a technological world change and whoever does not invest in science will receive a BAN

[Simargl]

Quote: viktor_ui

Artificial intelligence will make it possible to work at exorbitant aircraft overloads,

There is also a film not of the first freshness about an aircraft with AI.
There they began to use it as an escort aircraft (of the "Hunter" type).
So nothing new yet.

[Real Pilot]

And in the US, there has long been an idea to re-equip decommissioned combat aircraft with a control system, without a pilot. That is, to make autonomous attack vehicles, and with active air defense, these are kamikaze drones.
One F-35 carries a couple of F-16s, F-14s (decommissioned) or old F-15s as flying arsenals. It is more convenient than making new drones of this class.

Cheap, reliable, and even safer than a pure UAV. Many opponents are afraid to shoot at them, there are also American pilots (it seems). The Turks or their proxies (as they say) shot down 2 Rippers, but the F-16s would be ashamed!

[outsider]

- Within five years, experimental UAV fighters will appear ...

[Simargl]

What's wrong with "Hunter"?

[Pavel57]

The main question is what information was given by the AI ​​in the modeling process. Were there any restrictions on the field of view? The same limitation will also affect in real combat. Plus, AI potentially removes the restriction on aircraft overload, or, more precisely, allows it to increase, but not indefinitely.

[Blackmokona]

The F-35 has a 360-degree view, which is complete information for a dog duel.

[Pavel57]

BlackMokona Today,
The F-35 has a 360-degree view, which is complete information for a dog duel.
The F-35 has no eyes, so the question is, what kind of eyes was the AI ​​given in this simulation?

[voyaka uh]

Watch the video. At the end.
The battle was between F-16. The F-16 does not have IR video cameras around the glider perimeter,
like the F-35.
Therefore, the AI ​​was attached to the virtual helmet and joystick.
Exactly the same as that of a person. Ace pilot.
They had a level playing field.

[Blackmokona]

The F-35 has eyes and automatic target detection and identification.

[outsider]

- For AI, no one bothers to place electro-optical sensors not only within 360 ° in azimuth, but also within a full sphere, like the F-22 and even more so the F-35. By accepting information in this way, the AI ​​will have much better situational awareness than any pilot. Of course, the AI ​​will also receive the same data from the radar and the RTR station, such as the pilot on the F-35.

[Operator]

AI allows the plane to fly with a greater overload than a manned plane - due to that AI wins in a net (except for group battles, where tactics are important, in which the person is still strong).

[Pavel57]

AI in chess beats the human champion. At the same time, there are two directions of AI in chess - a set of algorithms that take into account the experience of people, and based on neural networks, which can learn itself.
Such an AI reached the level of a grandmaster in 4 hours of playing with itself.

[Operator]

In my comment, we were talking about the initial handicap of AI (regardless of its level of intelligence) - the ability to withstand extreme overloads and, accordingly, reduce the time it takes to perform aerobatics.

The human being has become the limiting factor in the pilot airplane complex.

[Danila46]

so Hto won ????

[iouris]

Quote: Danila46

so Hto won ????

But this is unknown. But the man lost.

[Evgesha]

Chrynovy pilot ..
He has few hours in the next games
Chuck Yager
f-19 (expert level)
f-117 (expert level)
Retaliator
he had to play a hundred hours in these games and hack to death,
when in RETAL he would have learned to land a fighter on a deck
then he could go out to battle with the computer.
and since the sucker was done.

[stroybat ZABVO]

At 7 hi.
If (yes, sort of) instead of a "pilot" there was a "gamer", God forgive me, then he WOULD make a five-star hero of the United States out of AI, tore the AI ​​like a monkey tore apart a newspaper Washington, ugh, potz ...

[tuts]

no, they just wrote the pilots of military aviation in the United States is not enough

[Shahno]

It's actually about this. Cheers up ...

[Kuzmich Sibiryakov]

EGESNIKI rule.
This is to what extent you need not have a head to do such things, and even excitedly write about it!
Is it really not clear that any virtual actions are not human actions, but the action of a program. TOTAL, the programs "fought" against each other. One is operatively controlled by a person, the second is working according to its own algorithm. A training battle should take place IN NATURE, on a range, with real combat units, and with imitation of shooting, well, at least towards a real pilot. In the direction of the drone, it is better, of course, to fire off real supplies. This will be transparent, evidence-based and indisputable.