A.I. Downs Expert Human Fighter Pilot In Dogfights

New and old developments in aviation technology.
  • Author
  • Message
Offline
User avatar

KamenRiderBlade

Elite 2K

Elite 2K

  • Posts: 2635
  • Joined: 24 Nov 2012, 02:20
  • Location: USA

Unread post28 Jun 2016, 05:46

http://www.popsci.com/ai-pilot-beats-ai ... SOC&dom=tw

We need these types of AI in cheap bug high capability drones to back up our human pilots.

The team work that they can do would allow them to dominate any situation.
Offline

35_aoa

Forum Veteran

Forum Veteran

  • Posts: 506
  • Joined: 28 Apr 2015, 04:03
  • Location: Virginia Beach, VA

Unread post28 Jun 2016, 08:41

Hold on for a second. The article claims the guy flying as the human is an Air Battle Manager. That is an AWACS controller……i.e. not a fighter pilot, not even a pilot. Colonel or not, he has about as much "dogfighting" experience as my 11 month old son.
Offline
User avatar

KamenRiderBlade

Elite 2K

Elite 2K

  • Posts: 2635
  • Joined: 24 Nov 2012, 02:20
  • Location: USA

Unread post28 Jun 2016, 10:04

And "Geno" is no slouch. He's a former Air Force Battle Manager and adversary tactics instructor. He's controlled or flown in thousands of air-to-air intercepts as mission commander or pilot. In short, the guy knows what he's doing. Plus he's been fighting A.I. opponents in flight simulators for decades.
Offline

neurotech

Elite 2K

Elite 2K

  • Posts: 2346
  • Joined: 09 May 2012, 21:34

Unread post29 Jun 2016, 02:07

35_aoa wrote:Hold on for a second. The article claims the guy flying as the human is an Air Battle Manager. That is an AWACS controller……i.e. not a fighter pilot, not even a pilot. Colonel or not, he has about as much "dogfighting" experience as my 11 month old son.

Don't you have model F/A-18s and MiG-29s for him to play with? Start teaching BFM skills when he's young...

Something about this whole story doesn't make any sense, because this AI controlled drone apparently can perform maneuvers beyond a human pilot can handle, that when flown in a F-22 would cause the jet would be a museum piece afterwards. 91-4003 Is there any indication that this simulator was programmed with the actual kinematics of the aircraft? Missile Pk? There have been bandits who have evaded multiple missiles before being shot down, so this AI drone has how many AIM-9s and AIM-120s?

Wasn't there an Israeli fighter pilot, Giora Epstein, who evaded all the bandits in a 11v1 dogfight? Shooting down at least 3 MiG-21s. I wonder how the AI would handle an adversary like that? Then again Col. Epstein is wrong about one thing. The F-16 doesn't fly itself in a dogfight. It still takes a pilot to fly it.
Offline

madrat

Elite 2K

Elite 2K

  • Posts: 2220
  • Joined: 03 Mar 2010, 03:12

Unread post29 Jun 2016, 03:12

Sometimes I don't think you read the article, neurotech. The AI was handicapped to the point it wasn't anywhere near a match physically. The AI was even given handicapped weapons and rules of engagement. Point is it used attrition to wear down the human.
Offline

35_aoa

Forum Veteran

Forum Veteran

  • Posts: 506
  • Joined: 28 Apr 2015, 04:03
  • Location: Virginia Beach, VA

Unread post29 Jun 2016, 03:30

KamenRiderBlade wrote:
And "Geno" is no slouch. He's a former Air Force Battle Manager and adversary tactics instructor. He's controlled or flown in thousands of air-to-air intercepts as mission commander or pilot. In short, the guy knows what he's doing. Plus he's been fighting A.I. opponents in flight simulators for decades.


Not sure if quoting just to back up what I said, or if you are pointing out that they said he flew intercepts as pilot. That much is certainly not true, based both on his career as an ABM, as well as his LinkdIn page (yeah I'm a creeper). Standard media mis-information.
Offline

neurotech

Elite 2K

Elite 2K

  • Posts: 2346
  • Joined: 09 May 2012, 21:34

Unread post29 Jun 2016, 03:37

madrat wrote:Sometimes I don't think you read the article, neurotech. The AI was handicapped to the point it wasn't anywhere near a match physically. The AI was even given handicapped weapons and rules of engagement. Point is it used attrition to wear down the human.

Where in the article does it say the AI was handicapped? (Not talking about the original paper)

The part about superhuman G limits was based on a comment on another site, for the the same story.

There are several versions of the story on different sites, and none seem to make sense as a conclusion, or they leave details out. The AI being handicapped physically may be the case, but I still think the conclusion as presented in the article is not based on the reality of a dogfight.
Offline

basher54321

Elite 1K

Elite 1K

  • Posts: 1718
  • Joined: 02 Feb 2014, 15:43

Unread post29 Jun 2016, 20:01

neurotech wrote:There are several versions of the story on different sites, and none seem to make sense as a conclusion, or they leave details out. The AI being handicapped physically may be the case, but I still think the conclusion as presented in the article is not based on the reality of a dogfight.


Agree with that - little to no information in them really and got to say it doesn't sound that impressive going by what is there anyway.
Offline

arian

Banned

  • Posts: 1293
  • Joined: 23 Dec 2014, 09:25

Unread post29 Jun 2016, 21:15

This is the exact quote form the journal article, not the condensed journalistic version:

As a former United States Air Force Air Battle Manager, Geno
is a United States Air Force Fighter Weapon School graduate and
Adversary Tactics (Aggressor) Instructor, and has controlled or flown
in thousands of air-to-air intercepts as a Ground Control Intercept
officer, as a Mission Commander on AWACS, and in the cockpit of
multiple fighter aircraft.


So it doesn't say that he himself has flown thousands of air-air missions.

Second, what his qualifications as a pilot are is irrelevant. He helped develop the AI he is testing. So being an expert in tactics, why is it surprising that an AI designed specifically to know all that these experts know, would be a strong opponent? That's what they were trying to design.

Geno noted how the first generation of red ALPHA held its own
against the blue variant of ALPHA, but the resulting engagements often
ended with heavy losses for both sides. Psibernetix and Geno worked
together to develop tactics, techniques, and procedures to overcome
red ALPHA’s payload and no-AWACS disadvantage, capitalize on
blue’s mistakes, and take advantage of numeric platform superiority
(when the situation presented itself). The current revised red ALPHA
model presented the blue adversary with credible offensive tactics
and timely defensive reactions that challenged blue’s radar sort logic,
compressed their engagement timeline, and rapidly put blue into a
defensive position from which they could not escape. The net result
after a prolonged engagement was blue’s total defeat with no or
minimal losses by red ALPHA.
When Geno took manual control of the blue aircraft against
the reds controlled by the baseline controller AFRL had previously
been utilizing, he could easily defeat it. However, even after repeated
attempts against the more mature version of ALPHA, not only could
he not score a kill against it, he was shot out of the air by the reds every
time after protracted engagements. He described ALPHA as “the most
aggressive, responsive, dynamic and credible AI (he’s) seen-to-date.”


Geno was the guy who trained the AI. An AI designed for tactics, not as a realistic pilot simulator. A fact that Popular Science apparently forgot to mention in order to make it look like some random student developed an AI, and then went out and beat a pilot with it.

There's no dog-fighting involved here. This is a stand-off tactics simulator, and their intent here is not to realistically simulate every aspect of blue vs. red capabilities. They just want to figure out how to develop algorithms that can more efficiently make decisions.
Offline

neurotech

Elite 2K

Elite 2K

  • Posts: 2346
  • Joined: 09 May 2012, 21:34

Unread post29 Jun 2016, 22:21

arian wrote:This is the exact quote form the journal article, not the condensed journalistic version:

As a former United States Air Force Air Battle Manager, Geno
is a United States Air Force Fighter Weapon School graduate and
Adversary Tactics (Aggressor) Instructor, and has controlled or flown
in thousands of air-to-air intercepts as a Ground Control Intercept
officer, as a Mission Commander on AWACS, and in the cockpit of
multiple fighter aircraft.


So it doesn't say that he himself has flown thousands of air-air missions.

Second, what his qualifications as a pilot are is irrelevant. He helped develop the AI he is testing. So being an expert in tactics, why is it surprising that an AI designed specifically to know all that these experts know, would be a strong opponent? That's what they were trying to design.

What his qualifications are as a pilot do matter if he's the one doing the flying.

Secondly, when was the last time a real dogfight stayed BVR with bandits still in play? Its called "the merge" for a reason. Classic example would be 1989 Gulf of Sidra incident. During the engagement, the F-14s tried to maintain an offset course, but when that didn't work, two AIM-7s were fired at the MiG, and missed. The second F-14 shot down one MiG-23 at 5 miles with an AIM-7, and then the lead F-14 closed for an Sidewinder kill.

https://en.wikipedia.org/wiki/Gulf_of_S ... dent_(1989)

It sounds suspiciously like this guy wasn't familiar with the actual flying techniques to counter an incoming missile. Another detail is the simulator might not have accurately represented the incoming missile to the pilot.
Offline

arian

Banned

  • Posts: 1293
  • Joined: 23 Dec 2014, 09:25

Unread post29 Jun 2016, 23:09

neurotech wrote:
arian wrote:This is the exact quote form the journal article, not the condensed journalistic version:

As a former United States Air Force Air Battle Manager, Geno
is a United States Air Force Fighter Weapon School graduate and
Adversary Tactics (Aggressor) Instructor, and has controlled or flown
in thousands of air-to-air intercepts as a Ground Control Intercept
officer, as a Mission Commander on AWACS, and in the cockpit of
multiple fighter aircraft.


So it doesn't say that he himself has flown thousands of air-air missions.

Second, what his qualifications as a pilot are is irrelevant. He helped develop the AI he is testing. So being an expert in tactics, why is it surprising that an AI designed specifically to know all that these experts know, would be a strong opponent? That's what they were trying to design.

What his qualifications are as a pilot do matter if he's the one doing the flying.

Secondly, when was the last time a real dogfight stayed BVR with bandits still in play? Its called "the merge" for a reason. Classic example would be 1989 Gulf of Sidra incident. During the engagement, the F-14s tried to maintain an offset course, but when that didn't work, two AIM-7s were fired at the MiG, and missed. The second F-14 shot down one MiG-23 at 5 miles with an AIM-7, and then the lead F-14 closed for an Sidewinder kill.

https://en.wikipedia.org/wiki/Gulf_of_S ... dent_(1989)

It sounds suspiciously like this guy wasn't familiar with the actual flying techniques to counter an incoming missile. Another detail is the simulator might not have accurately represented the incoming missile to the pilot.


You can read the actual journal article for more details: http://www.omicsgroup.org/journals/gene ... 000144.pdf

This isn't a dogfight simulator, nor is it intended to be a realistic simulation of all technologies and capabilities involved. The baseline assumptions of the model are basic: missile range, missile pk, etc. It says right in the article they don't have a realistic 3d performance model for missiles yet etc. And it says in the article this is not at dog-fighting ranges.

But that's not the point of the AI simulation, contrary to what the Popular Science article is hinting at, or the way people are discussing it. The point of the simulation is to develop an AI that can take into account a large set of variables and make decisions in a more efficient way than previous AIs.

So the AI is beating other AIs, and humans, not by being super-human in physical capabilities or "dodging" missiles etc. It's about tactics, and the AI makes decisions on what tactics to employ faster and more efficiently than humans.

As for his qualifications, he instructs pilots on tactics, so I'm assuming he's quite good at it. But then again, he also trained the AI.
Offline

neurotech

Elite 2K

Elite 2K

  • Posts: 2346
  • Joined: 09 May 2012, 21:34

Unread post30 Jun 2016, 00:32

arian wrote:As for his qualifications, he instructs pilots on tactics, so I'm assuming he's quite good at it. But then again, he also trained the AI.

Do you know him? Have you seen him actually fly a fighter jet?

He may know tactics, but doesn't mean he has flying skills representative of a fighter pilot. At both USAF WS and TOPGUN, they have Subject Matter Experts as weapons and tactics instructors the ground, but only fighter pilots get to fly a fighter jet as instructors. Its like comparing an Air Traffic Controller to an Airline Captain, both are highly skilled, but have completely different roles.

There was a time when the "experts" said that dogfighting was obsolete, and long-range missiles were the future. They were wrong. The Navy created TOPGUN to teach dogfighting skills to fighter pilots. AI or not, the idea that air-to-air can be fought exclusively at "standoff" range is stupid.

The original research paper presented an implementation of a machine learning algorithm based on a contrived scenario that has little basis in reality.

I'm not a TOPGUN pilot, but I do work as a programmer on Machine Learning applications. Like most media stories based on papers with anything to do with machine learning, neural networks or AI related subjects, the media story is usually b##sh#t!
Offline
User avatar

popcorn

Elite 5K

Elite 5K

  • Posts: 7690
  • Joined: 24 Sep 2008, 08:55

Unread post30 Jun 2016, 00:41

More on Cyberdyne.. oops I mean Psibernetix :D
http://www.psibernetix.com/projects/defense/
"When a fifth-generation fighter meets a fourth-generation fighter—the [latter] dies,”
CSAF Gen. Mark Welsh
Offline

arian

Banned

  • Posts: 1293
  • Joined: 23 Dec 2014, 09:25

Unread post30 Jun 2016, 01:22

neurotech wrote:
arian wrote:As for his qualifications, he instructs pilots on tactics, so I'm assuming he's quite good at it. But then again, he also trained the AI.

Do you know him? Have you seen him actually fly a fighter jet?

He may know tactics, but doesn't mean he has flying skills representative of a fighter pilot. At both USAF WS and TOPGUN, they have Subject Matter Experts as weapons and tactics instructors the ground, but only fighter pilots get to fly a fighter jet as instructors. Its like comparing an Air Traffic Controller to an Airline Captain, both are highly skilled, but have completely different roles.

There was a time when the "experts" said that dogfighting was obsolete, and long-range missiles were the future. They were wrong. The Navy created TOPGUN to teach dogfighting skills to fighter pilots. AI or not, the idea that air-to-air can be fought exclusively at "standoff" range is stupid.

The original research paper presented an implementation of a machine learning algorithm based on a contrived scenario that has little basis in reality.

I'm not a TOPGUN pilot, but I do work as a programmer on Machine Learning applications. Like most media stories based on papers with anything to do with machine learning, neural networks or AI related subjects, the media story is usually b##sh#t!


That may well be the case, but it's tangential to the question. The media version of the story is obviously sensational and wrong.
Offline

gideonic

Senior member

Senior member

  • Posts: 280
  • Joined: 06 Sep 2015, 13:54

Unread post01 Jul 2016, 12:36

arian wrote:You can read the actual journal article for more details: http://www.omicsgroup.org/journals/gene ... 000144.pdf

This isn't a dogfight simulator, nor is it intended to be a realistic simulation of all technologies and capabilities involved. The baseline assumptions of the model are basic: missile range, missile pk, etc. It says right in the article they don't have a realistic 3d performance model for missiles yet etc. And it says in the article this is not at dog-fighting ranges.


Yes good point. I also really really suggest you read the actual article, before bashing it. Here is a good condensed version: https://www.reddit.com/r/CredibleDefens ... zzy_based/

ALPHA’s [the AI's] current primary objective is to serve as an intelligent hostile force for pilots to train against within the AFSIM simulation environment [8]. While the models showcased within this document for platforms do not mimic any system specifically, AFSIM realistically represents a modern air combat environment with appropriately behaving models for aircraft, sensors, and weapons. Data is collected from sensors, fused real-time, and sent to ALPHA, complete with noise and potential failures. This data feed is received at and commands sent to unique UDP ports for each platform, causing this problem to contain some of the software complexities that would be present in actual hardware implementation.

For current mission profiles, ALPHA’s red forces are handicapped with shorter range missiles and a reduced missile payload than the blue opposing forces. ALPHA also does not have airborne warning and control system (AWACS) support providing 360° long range radar coverage of the area; while blue does have AWACS. The aircraft for both teams are identical in terms of their mechanical performance. While ALPHA has detailed knowledge of its own systems, it is given limited intelligence of the blue force a priori and must rely on its organic sensors for situational awareness (SA) of the blue force; even the number of hostile forces is not given. Both to mirror training exercises and to offset these weaknesses, ALPHA is typically given a numeric advantage over the blue forces. However this is not always the case and ALPHA is capable of controlling any finite number of friendly aircraft. The current problem is focused on purely beyond visual range air-to-air combat missions; no ground targets or friendly platforms requiring escort are considered at this time.

The mission analyzed in this document features two blue fighters vs. four red [controlled by ALPHA]. The red aircraft begin over a defended coastline and the blues are 54 nautical miles due west. The blues each have 4 long range missiles (LRMs) and 4 acrobatic short range missiles (SRMs), whereas the reds have 4 medium range missiles (MRMs) onboard each platform. The reds’ radar is long range with a +/- 70° azimuth angle and a 15 degree elevation angle. The initial state of the mission is displayed below in Figure 1, with the blue AWACS off-screen, due northwest of the blue fighters.

Again, ALPHA is currently trained to utilize a force of superior numbers, but weaker capabilities. In order to complete missions safely, the blue forces must be manipulated into poor positions, as their kill range is farther than the reds’.

The paper also has nice pictures of these many-to-many scenarios, I highly recommend taking a look. This was no dumbed down 1-1 dogfight simulator. And yes, they had force superiority, but still it was no cakewalk (poorer missiles, no AWACS). The pilot was easily able to "win" the same scenario against the baseline AI.
Next

Return to Technology

Who is online

Users browsing this forum: No registered users and 3 guests