Unfortunate AI Behavior

[PatrickAWlson]

Since I have switched to 109s and taking on more fighters, I have seen some AI behavior that is less than great.  

 

1. One of my flight mates attacking IL2s from an altitude advantage.  Ends up screaming into the ground near vertical.  Flying a 109 F4.  Guessing he got too fast and controls locked.

2. AI seems to overdo it and black out frequently.  At least four crashes of flyable planes, generally during a maneuvering contest.  These are not necessarily lawn darts.  They look like a plane with an incapacitated pilot.

3. Damaged IL2 breaks off the flight and then goes aerobatic.  Seen this twice.  They start doing things in a near dead aircraft that would be tough to replicate in an undamaged one.

 

Overall, kudos to the team for the AI improvements that have been made to date.  The AI really has made great strides.  Just pointing out some of the issues that I have seen.

 

Request:

1. Certain aircraft types should never go full aerobatic.  A badly damaged IL2, for instance, should never go vertical.  I have seen any number of planes seemingly use fighter AI that are not fighters (B24s, A20s, PE2s, etc.)

2. Might want to scale back the G forces an AI pilot is willing to take or generally pay better attention the way that humans have to.  This seems to be, by a long shot, the #1 cause of self induced demise.  Alternatively, maybe give AI and humans a bit more tolerance.  My guy seems to start sucking wind pretty quickly.

[Gambit21]

Can they shoot yet?

My AI vs AI fur-balls were taking damn near 20 minutes to resolve.

 

[oc2209]

3 hours ago, Gambit21 said:

Can they shoot yet?

My AI vs AI fur-balls were taking damn near 20 minutes to resolve.

 

 

Depends on the armament type and configuration (wing versus nose) and a few other factors that are difficult to quantify. In other words, the AI is deadly accurate sometimes, and other times it's wildly inaccurate. Obviously the latter more than the former. But I have seen AI-controlled 109Fs take down Sturmoviks without wing guns, which is pretty damn impressive.

 

I don't really consider it a failing of the AI to not wipe the skies clean in 5 minute dogfights. Even considering that yes, on Ace, it should theoretically be able to do so. In reality that level of aggression and accuracy would be simply intolerable for humans to compete with and/or enjoy fighting against. 

 

The AI seems unable to hit with large cannons (30mm+) the vast majority of the time, and it generally fails to pull enough lead against turning targets (using any guns). At most, it can hit the target's tail in turns.

 

Having said all that, I feel like aiming is the least of the AI's worries at the moment. It's accurate enough to be a threat a decent percentage of the time. That's all it really needs to be. Where advancements need to be made, are in the realm of situational awareness (or the illusion thereof) and self-preservation behaviors.

 

Take, for instance, the lawn darting phenomenon. Perhaps a simple way to fix it would be to have the AI pull out of any dives once it reaches a certain speed below a certain altitude. This would prevent it from making high-speed diving attacks against low-flying targets. Like, it will automatically level out when flying over 300 MPH at 1,000 feet altitude or less. It will stop diving above 325 MPH at 2,000 feet. 350 MPH at 3,000 feet. This way you don't have to program anything more complex (like dive angles, acceleration, G-forces, etc) than speed and altitude awareness.

 

As for other things @PatrickAWlson mentioned, I'm not sure how you could tell the AI to fly conservatively when damaged. One of the problems I've observed lately with both the Stuka and Sturmovik is that they both have such large wing areas, they both seemingly shrug off severe aerodynamic damage.

 

I think the blackout issue could be pretty easily solved by just having the AI ease off its turns at an earlier point in its G-fatigue level. I have seen it tire itself out after, say, 5 minutes of turning, whereupon in noticeably turns more loosely. So, instead of letting it reach that point of near-zero endurance left, make the cutoff higher, at, say, 25% of its total endurance, or even as high as 50%. This way it always has something left in its reserves, and is less likely to accidentally black out.

 

The downside is that it will be undeniably easier to shoot down in long turning engagements. But if that's preferable to consistent blacking out and crashes, so be it.

 

Maybe the G-fatigue logic could also be tied in with altitude. Example: at low altitudes (under 2,000 feet, say) the AI is more cautious with its G-endurance, and eases out of tight turns when it reaches 50% fatigue. At higher altitudes (thus with a higher probability of regaining consciousness before hitting the ground), the AI will spend more of its G-endurance, up to the very limit (zero fatigue remaining).

[PatrickAWlson]

10 hours ago, oc2209 said:

 

As for other things @PatrickAWlson mentioned, I'm not sure how you could tell the AI to fly conservatively when damaged. One of the problems I've observed lately with both the Stuka and Sturmovik is that they both have such large wing areas, they both seemingly shrug off severe aerodynamic damage.

 

I think the blackout issue could be pretty easily solved by just having the AI ease off its turns at an earlier point in its G-fatigue level. I have seen it tire itself out after, say, 5 minutes of turning, whereupon in noticeably turns more loosely. So, instead of letting it reach that point of near-zero endurance left, make the cutoff higher, at, say, 25% of its total endurance, or even as high as 50%. This way it always has something left in its reserves, and is less likely to accidentally black out.

 

The downside is that it will be undeniably easier to shoot down in long turning engagements. But if that's preferable to consistent blacking out and crashes, so be it.

 

Maybe the G-fatigue logic could also be tied in with altitude. Example: at low altitudes (under 2,000 feet, say) the AI is more cautious with its G-endurance, and eases out of tight turns when it reaches 50% fatigue. At higher altitudes (thus with a higher probability of regaining consciousness before hitting the ground), the AI will spend more of its G-endurance, up to the very limit (zero fatigue remaining).

 

For many non fighter plane types I would like to see restrictions in the types of maneuvers.  B25s should not bank at 90 degrees.  IL2s should not go vertical.   I could envision code where restrictions were encapsulated in a piece of code and then configured for the different planes.  Fighters would not have a roll restriction where bombers might be limited to 30 degrees.  Speed/dive/altitude considerations could be modeled as an algorithm (pull up/throttle back) with different parameters for different planes.  

 

For blackout, the AI is flying under the same penalties as humans so there must be some awareness in the code.  As you point out, better to insist that your opponent shoot you down than to crash for him.

[Thad]

14 hours ago, PatrickAWlson said:

Since I have switched to 109s and taking on more fighters, I have seen some AI behavior that is less than great.  

 

1. One of my flight mates attacking IL2s from an altitude advantage.  Ends up screaming into the ground near vertical.  Flying a 109 F4.  Guessing he got too fast and controls locked.

2. AI seems to overdo it and black out frequently.  At least four crashes of flyable planes, generally during a maneuvering contest.  These are not necessarily lawn darts.  They look like a plane with an incapacitated pilot.

3. Damaged IL2 breaks off the flight and then goes aerobatic.  Seen this twice.  They start doing things in a near dead aircraft that would be tough to replicate in an undamaged one.

 

Overall, kudos to the team for the AI improvements that have been made to date.  The AI really has made great strides.  Just pointing out some of the issues that I have seen.

 

Request:

1. Certain aircraft types should never go full aerobatic.  A badly damaged IL2, for instance, should never go vertical.  I have seen any number of planes seemingly use fighter AI that are not fighters (B24s, A20s, PE2s, etc.)

2. Might want to scale back the G forces an AI pilot is willing to take or generally pay better attention the way that humans have to.  This seems to be, by a long shot, the #1 cause of self induced demise.  Alternatively, maybe give AI and humans a bit more tolerance.  My guy seems to start sucking wind pretty quickly.

 

Heck, at my age, I get winded climbing into the cockpits. ?

Edited August 8, 2021 by Thad

[PatrickAWlson]

17 minutes ago, Thad said:

 

Heck, at my age, I get winded climbing into the cockpits. ?

 

That's why I play flight sims and RPGs.  Forever young.  On the computer.  As long as I have icons to help me see things. And difficulty settings that I can drop to easy when my no longer 15 year old reflexes are not up to it.  

[oc2209]

5 hours ago, PatrickAWlson said:

 

For many non fighter plane types I would like to see restrictions in the types of maneuvers.  B25s should not bank at 90 degrees.  IL2s should not go vertical.   I could envision code where restrictions were encapsulated in a piece of code and then configured for the different planes.  Fighters would not have a roll restriction where bombers might be limited to 30 degrees.  Speed/dive/altitude considerations could be modeled as an algorithm (pull up/throttle back) with different parameters for different planes.

 

I think a roll restriction is the wrong way to go about it, because it's a rather arbitrary limitation. Rather, it should be measured by G-forces. Since bombers naturally have lower tolerance levels for hard maneuvers, make the AI avoid pulling anything over 2.5 G in a bomber. At least not for any sustained amount of time.

 

With IL-2s going vertical, I don't see an easy way to prevent it. And I'm not sure I want it prevented. Sure, at face value it looks ridiculous. But since such things are possible (it's easier to be safely aerobatic in a Stuka/Sturmovik than a 2-engine bomber), I think allowances should be made for exceptionally talented attack plane pilots. Everyone shouldn't have to be a sitting duck just because they're flying a slug. As long as the physics holds up (as in, you can try to zoom climb, but you won't get far, and you'll stall down twice as much altitude as you gained), I'm all for unconventional flying. So long as all the AI attack pilots aren't barnstormers. Most will still plod along to their doom. A few will try to get creative.

 

*Edit*

 

I just did a few test flights, and multi-purpose bombers like the Ju-88 and Pe-2 have very high G-limits. There is no valid reason why they can't try to maneuver the same way a Bf-110 flies, for instance. They'll simply be less capable than the 110.

 

Level bombers like the He-111 (and presumably the B-25?) have lower G-limits. The A-20 is in between the Ju-88 and He-111. Transport planes have the lowest tolerances. However, that said, I still had a hard time exceeding 2.5 G in a tight turn. I had to reach 300 MPH in a dive to rip the wings off my He-111 in the subsequent pull up. Most of the time, under 300 MPH, a twin engine bomber will be unable to generate much G-force even in radical maneuvers. Looping a Ju-52 from a ~215 MPH dive will barely bring the G-load to 4.0, and only then for a few seconds. So it's the speed, not the angles of certain maneuvers, that apparently stresses the airframe most.

 

That being the case, I don't see a convincing reason an He-111 travelling at ~200 MPH can't or shouldn't enter a turn with a 90-degree bank. The only argument against such maneuvers would be that conventional wisdom says bombers don't fly that way. But that could just be the limits of bomber pilot training more than physical plane limits. Another factor could be that stick/wheel forces are much less for us and AI pilots than they would be for actual bomber pilots. Since, however, stick forces aren't modelled for us, they shouldn't be modelled for AI pilots either.

Edited August 8, 2021 by oc2209

[Mad_Mikhael]

On 8/8/2021 at 5:02 AM, PatrickAWlson said:

Since I have switched to 109s and taking on more fighters, I have seen some AI behavior that is less than great.  

 

1. One of my flight mates attacking IL2s from an altitude advantage.  Ends up screaming into the ground near vertical.  Flying a 109 F4.  Guessing he got too fast and controls locked.

 

I've seen this a lot on the Finnish server.

Looks like we've achieved skills level of AI same as humans. Good job devs!

 

 

?

[Feathered_IV]

On 8/8/2021 at 1:36 PM, Gambit21 said:

Can they shoot yet?

My AI vs AI fur-balls were taking damn near 20 minutes to resolve.

 

 

I'm finding that too.  8 vs 8 QMBs with Ace AI and I'm shooting down 6 or 7 while the other one goes home.  Neither side manages score against their AI counterparts very often though.  Perhaps I'm just that good.

[Yogiflight]

On 8/8/2021 at 9:59 PM, oc2209 said:

I don't see a convincing reason an He-111 travelling at ~200 MPH can't or shouldn't enter a turn with a 90-degree bank.

Maybe because there is a bit more crew onboard, than just the pilot. The front gunner, the ventral gunner and especially the waist gunner would be in serious troubles, when the pilot would fly the 111 the way you can see it in the game.

The problem here is, the AI can't fly turns with the rudder, like you can see it in IL-2 1946. In the old IL-2 all turns except in dogfights were flown with rudder, no the hard fighter turns, you can see even for small changes of heading in this game.

Take the 110. The gunner doesn't even have a backrest. If you would fly a real 110 the way it is done in the game all the time, the gunner would have a really hard time, to cling on anywhere in his compartment. Not to talk about reloading the 20mm guns in the E2 while the pilot is flying acrobatic and shooting at the same time, too.

[Stonehouse]

On the maneuver restrictions for bombers etc, not sure it should be by aircraft type or not solely by aircraft type. It is probably a pretty difficult thing to turn into an reliable algorithm.

 

Probably should be linked to type of maneuver and crew skill and perhaps even attacker type. Upwards vertical moves by bombers would probably not be common or very likely you'd think, particularly once damaged although I have read accounts by RNZAF P-40 pilots in the Pacific being regularly out climbed by B-24s after they have dropped their bomb load if the altitude was high enough (The Blue Arena - Bob Spurdle) so potentially against the right opponent and situation the B24 or other large aircraft might be able to use that technique. According to the book it gave the P-40s tasked with escorting the B-24s a real problem.

 

Another surprising example which probably people have read/heard of https://vwma.org.au/explore/units/1800.

 

I've also read of Lancasters and Sunderlands performing surprisingly violent defensive maneuvers although as I remember all downwards such as rolling corkscrew dives. Must have been very tough on some of the crew positions during such a maneuver but they still did it when needed. He111s and Ju87s had pretty bad climb rates as I remember so don't think they would tend to pull vertical climbs but reckon they could do very surprising diving moves if their survival was at stake. I don't think they would worry about over stressing the aircraft or bouncing the gunners about. I think the more confident and skilled the pilot was the more likely they would try odd moves to escape being shot down at all costs. The story of the Hudson and it's crew stands out I guess because they kept the fight going for a very long time in terms of air combat (around 10 mins for 1 Hudson v's 9 circa 1942 Zeroes) against a skilled group of pilots and repeatedly attempted to maneuver offensively. I don't know but guess a B25 or A20 or Pe-2 type aircraft with the right crew might be capable of the same type of thing. For the 110 I tend to expect it to fly like a fighter but that while doing such things the rear gunner should be much less accurate. Maybe it is the gunner accuracy during violent maneuvers at least for bombers that is the main problem?  

[[DBS]Browning]

On 8/8/2021 at 6:54 PM, PatrickAWlson said:

 

For many non fighter plane types I would like to see restrictions in the types of maneuvers.  B25s should not bank at 90 degrees.  IL2s should not go vertical.   I could envision code where restrictions were encapsulated in a piece of code and then configured for the different planes.  Fighters would not have a roll restriction where bombers might be limited to 30 degrees.  Speed/dive/altitude considerations could be modeled as an algorithm (pull up/throttle back) with different parameters for different planes.  

 

For blackout, the AI is flying under the same penalties as humans so there must be some awareness in the code.  As you point out, better to insist that your opponent shoot you down than to crash for him.

 

I don't know about this.

The ~45 degree bank "cockscrew" maneuver was standard practice for evasive Lancasters and all our bombers are considerably more maneuverable than that.

 

[unreasonable]

Good description of that here: https://masterbombercraig.wordpress.com/avro-lancaster-bomber/corkscrew-port-go/

 

Mind you, Lancasters etc doing this would be flying alone vs single night fighters and were all trained to do it: if you were trained to stay in formation at all costs, to maximise defence vs enemy fighter formations, I expect the range of manoeuvres the pilot would try would be very limited. So it is hard to generalise.  

[PatrickAWlson]

8 hours ago, [DBS]Browning said:

 

I don't know about this.

The ~45 degree bank "cockscrew" maneuver was standard practice for evasive Lancasters and all our bombers are considerably more maneuverable than that.

 

 

Totally agree that real maneuvers should not be abandoned.  I was referring more to the fighter maneuvers that bombers sometimes perform.  Bank angles well beyond 45 degrees.  Flight attitudes that include inverted.  Basically anything for which "hang on tight boys" would be woefully inadequate.  Especially if the bomber is one of a formation.

[AEthelraedUnraed]

On 9/23/2021 at 11:44 AM, [DBS]Browning said:

I don't know about this.

The ~45 degree bank "cockscrew" maneuver was standard practice for evasive Lancasters and all our bombers are considerably more maneuverable than that.

Evasive maneuvers are one thing, but some bombers (e.g. the recently released D.H.4) will go full fighter mode and agressively pursue the player, even if he flies away.

[[DBS]Browning]

2 hours ago, AEthelraedUnraed said:

Evasive maneuvers are one thing, but some bombers (e.g. the recently released D.H.4) will go full fighter mode and agressively pursue the player, even if he flies away.

I think the DH4 is a poor choice to illustrate this.

The forward gun in the DH.4 was not put there to strafe ground targets. Like the Bristol, and many other early twin-seats, it blurred the line between fighter and bomber and it is absolutely appropriate for this plane to be aggressive and acrobatic in combat. It's certainly more than capable of such things.

 

 

Edited September 24, 2021 by [DBS]Browning

[AEthelraedUnraed]

1 hour ago, [DBS]Browning said:

I think the DH4 is a poor choice to illustrate this.

The forward gun in the DH.4 was not put there to strafe ground targets, like the Bristol, and many other early twin-seats, it blurred the line between fighter and bomber and it is absolutely appropriate for this plane to be aggressive and acrobatic in combat. It's certainly more than capable of such things.

It *could*, but that doesn't mean it *always should*. If a large flight of bombers is attacked by a couple of fighters, I'd usually expect them to either:

- drop their bombs and run, staying in formation, or

- continue to the target and hope for the best, staying in formation.

 

Perhaps with an enemy on their tail, they'd do some funky maneuvering, but mostly I'd expect those planes to return to their formation once the immediate threat has disappeared. The DH.4 wikipedia page also seems to indicate that at least some squadrons remained in formation if enemy fighters were around:

Quote

55 Squadron developed tactics of flying in wedge formations, bombing on the leader's command and with the massed defensive fire of the formation deterring attacks by enemy fighters

 

Of course, things might have been different in artillery spotting / reconnaissance missions, or for certain agressive, talented pilots (there were a number of DH.4 aces after all). But as things stand in-game, DH.4s will completely forget about their target whenever an enemy appears, and actively engage it and even follow it if it flees. Ideally, this would be up to the mission designer, with the current behaviour perfectly fine if the target priority has been set to Low, and a more passive, formation based behaviour if the target priority is Medium (like the Stuka or IL2 behaviour). A target priority of High doesn't result in the DH.4 breaking formation, but this also disables the rear gunner. This effectively disables light bomber intercept missions for Flying Circus, since every single aircraft behaves like a fighter.

Edited September 24, 2021 by AEthelraedUnraed

[[DBS]Browning]

I certainly agree that both behaviors should be possible.