Getting the best out of the KI

[JG27_Steini]

Hello gentlemen,

 

i just started a new campaign in JG52 in Kuban area. I have some mission already done, but my AI comrades keep dying and dying. I fly agressive and shoot enemy fighter very quick but my comrades rarerly shoot or kill an enemy. At the end i kill 2-4 fighter and loose 1-3 squad member while in the same fight with AI. My difficulty settings are medium and i started with the lowest rank. Do you have any idea how to get more out of the AI? Will it make any difference on hard difficulty?

 

Thank you

[PatrickAWlson]

Focus on keeping them alive, even at the cost of a victory or two for you.  

[vanson77]

Accept that AI is crap and your wingmen are mostly flying like drunken monkeys, more obstacle than of any use. At least against enemy fighters. Against Bombers and IL-2´s they can be quite useful.
In PWCG (which i absolutely recomend) i started to send them home right at the start of the mission to save their lives, and in combat report i under-claim to give every wingman 1 victory until they become ace (5 kills). But even when ace they are pretty useless.

 

Thats one of the reasons i recently started to play CloD, the AI is so much better there. As purely offline-player the BoX-AI is almost game braking for me. Improvements were being made but so small and in such a slow pace... It really makes me sad because otherwise IL-2 BoX would be such a good game.

 

Sorry, this probably wont help you. Only suggestions i have is train them to become ace quickly, and keep the fights in mid and high altitude where AI seems to work a little better.

 

Edited August 5, 2020 by vanson77

[PatrickAWlson]

What @vanson77 said.  I found my AI mates were doing just fine in 109s once they got some experience.  In 190s they are next to useless, so plane type makes a big difference.

[easterling77]

i can only confirm what @vanson77 and espacially @PatrickAWIson are saying. Most sad for me is the fact, that my great love in this game is the 190 and it's so frustrating to see the AI turnfighting in 190s against yaks

 

I guess the AI can handle simpler manouvering like turnfights but advanced tactics in vertical manner - it's not able to.

So this might be one reason why the AI is sooooo much better in the 109 than in the 190.

Edited August 5, 2020 by easterling77

[RedKestrel]

2 minutes ago, easterling77 said:

i can only confirm what @vanson77 and espacially @PatrickAWIson are saying. Most sad for me is the fact, that my great love in this game is the 190 and it's so frustrating to see the AI turnfighting in 190s against yaks

To be fair, online you run into a lot of real humans trying to turnfight in 190s.

[easterling77]

4 minutes ago, RedKestrel said:

To be fair, online you run into a lot of real humans trying to turnfight in 190s.

OK didnt notice that because I'm a SP guy?

 

It's just sad to see that the AI can't use the potential of this aircraft - but i can understand that AI is a compolex and long going project

Edited August 5, 2020 by easterling77

[RedKestrel]

1 minute ago, easterling77 said:

OK didnt notice that because I'm a SP guy?

 

 

As a man guilty of attempting to turnfight in a P-47 on occasion...I am in no position to throw stones at anybody.

[vanson77]

Another thing that helps is to start you career as commander, so you are always flight leader. Then you have some control over your wingmen and you can abort the mission if needed, or at least command your comrades to follow you and get out of the fighting zone which can be better in some situations.
Be cautious and as predictive as possible, sometimes its better to run and avoid fights (and pray that they don´t follow you across the whole map, which still occurs).

 

Btw sending them home directly in a fight is quite dangerous, because they fly home in a straight way ignoring everything at their tail, even when an enemy fighter is emptying his magazine at him.

 

[JG27_Steini]

Thx for all your response. Very frustrating that the game not really developed in this direction. So many SP functions are not working. After all this years this game is in SP far behind old IL2 standards with  bad AI, missing radio system, formation flying etc. 

 

They just keep building new maps with new planes instead of repair basic functionality.

[vanson77]

Jep. But new content sells, improvements of existing flaws not so much.

But who knows, maybe we get a "Working AI" DLC someday 

[PatrickAWlson]

Giving the team credit, there have been very significant AI improvements over the past year.  With that in mind I am not complaining about the effort, just noting that for speed oriented planes there is still some work to do.  IMHO the biggest single improvement the team could make for the AI is adding logic for when NOT to be aggressive.  That is not an easy task.

[Sybreed]

your friendly AI squaddies are good at setting you up for kills. They usually all have an enemy plane flying behind them. For some reason, the enemy AI always seem more proficient than your friendlies.

 

And btw, making the AI use the properties of each aircraft properly would probably mean writing a different AI for each and different plane... doable, a but a huge undertaking. Right now, AI is very comfortable with ally planes, like Spitfire, Yaks and whatnot.

[PatrickAWlson]

55 minutes ago, Sybreed said:

your friendly AI squaddies are good at setting you up for kills. They usually all have an enemy plane flying behind them. For some reason, the enemy AI always seem more proficient than your friendlies.

 

And btw, making the AI use the properties of each aircraft properly would probably mean writing a different AI for each and different plane... doable, a but a huge undertaking. Right now, AI is very comfortable with ally planes, like Spitfire, Yaks and whatnot.

 

To some extent there already is different AI for each plane as every plane has its own unique requirements just to fly it.  However, for what I am referring to, it would be a different AI for different situations.  So instead of 40 different AI routines you would instead have maybe 2 or 3 or 4.  The AI would choose which one to use based on its plane in the simplest case, and maybe also considering the opponents plane. 

 

To their credit they are already doing that.  Many AI improvements have been made around faster B&Z style fighting.  Now do a bit more.  Make the AI break off and refuse to engage in a circling fight.  Make the AI understand that low and slow is bad in many plane types and break off before it runs out of altitude and options.  Small, incremental steps.

[oc2209]

4 hours ago, PatrickAWlson said:

To their credit they are already doing that.  Many AI improvements have been made around faster B&Z style fighting.  Now do a bit more.  Make the AI break off and refuse to engage in a circling fight.  Make the AI understand that low and slow is bad in many plane types and break off before it runs out of altitude and options.  Small, incremental steps.

 

I do tons of quick battle 1v1s just for fun, all sorts of plane combinations (including pitting Allied against Allied and German against German).

 

I've noticed that the AI already exhibits behavior similar to what you want: breaking off and refusing to engage in a turn fight. The determining factor is the plane you're in. When, for example, I fly in a 109F-4 versus a P-47 (that I cruelly hobble by putting full fuel, extra ammo, and rockets on the wings; all at 300m starting altitude), the P-47 (on ace, at least) will attempt to fly on ahead of me and outrun me in a full-throttle, shallow climb. And even with all the extra weight, it will outrun me (I don't bother to use my emergency power for the 1 minute allowance). The only chance I've got, is to pull a very tight turn as we approach head on (from the battle's start positions), timed so that I come out of the turn directly behind him; I then have to fire at a pretty extreme range with only my 20mm, which usually cripples his engine enough that I can catch up to him. 

 

However, when I'm in a faster 109 version, he can't outrun me, and will always attempt to turn. At 300m, there's really not much else he can do. I thus use his predictability to test ballistics and deflection shooting.

 

There are other examples I have where P-38s try to 'lure' me up to higher altitudes by refusing to turn on the deck.

 

All that aside, I would say the AI's greatest weaknesses are 2 critical areas:

 

1) An inability to break once fired upon. A roll, a dive, a jink--anything would be preferable to no action at all, which is what I've seen a lot of AI wingmen do when fired on. Forget making the AI smart enough to not turn in a 109 when a Spitfire or Yak-9's behind them. Just do something.

 

2) Deflection shooting. The AI, as far as I can tell, cannot shoot with any kind of deflection. It will fire (inaccurately) in head-on passes, and will freely fire from directly behind or slightly above and directly behind. This is problematic for obvious reasons, chief among them: it teaches me extremely bad dogfighting habits, as I can routinely turn in front of an AI with a perfect position to fire, and it lets the opportunity pass. Consequently, I've never learned to improve my timing in a situation, for example, when you need to pick just the right moment to break when you spot a bounce setting up behind you. Breaking too early is almost as bad as not breaking at all, as I learned the hard way in multiplayer.

 

I think a better AI would 'train' people more for multiplayer; that's one of the best things it could do for the longevity/popularity of the game overall. Secondarily, it'd of course make single player more challenging and rewarding. Which would, again, probably lead to more sales by word of mouth and reviews. So in my opinion, it absolutely is worth investing in.

 

*edit: just as a case in point, I actually put off buying the game for years, because I heard the AI was weaker than 1946's. At this point, I'd rather the AI have unfair advantages and impeccable aiming, because it would, again, teach you to be a better pilot. The current AI teaches us nothing, beyond reinforcing bad habits.

Edited August 5, 2020 by oc2209

[Yogiflight]

1 hour ago, oc2209 said:

I do tons of quick battle 1v1s just for fun, all sorts of plane combinations

You can't compare AI behaviour in QMB with its behaviour in career and campaign missions, because it is completely different, what AI does.

 

1 hour ago, oc2209 said:

I've noticed that the AI already exhibits behavior similar to what you want: breaking off and refusing to engage in a turn fight. The determining factor is the plane you're in. When, for example, I fly in a 109F-4 versus a P-47 (that I cruelly hobble by putting full fuel, extra ammo, and rockets on the wings; all at 300m starting altitude), the P-47 (on ace, at least) will attempt to fly on ahead of me and outrun me in a full-throttle, shallow climb.

I experienced this kind of behaviour, too in QMB, flying a A8 and was outrun by the P 38, I wanted to fight with.

 

1 hour ago, oc2209 said:

Deflection shooting. The AI, as far as I can tell, cannot shoot with any kind of deflection. It will fire (inaccurately) in head-on passes, and will freely fire from directly behind or slightly above and directly behind.

This was changed some time ago. I can remember very well, how deadly AI fighters were in deflection shooting. They were hitting always in every angle, so you better did everything to not give them a chance to give you a deflection shooting, because then you were dead.

[Feathered_IV]

8 hours ago, JG27_Steini said:

Thx for all your response. Very frustrating that the game not really developed in this direction. So many SP functions are not working. After all this years this game is in SP far behind old IL2 standards with  bad AI, missing radio system, formation flying etc. 

 

They just keep building new maps with new planes instead of repair basic functionality.

 

Once the needs of every other real or imagined fringe group has been explored in the hope of a shortcut to commercial success they will eventually work their way back around to recognising where the vast majority of their client base is.  

[BraveSirRobin]

43 minutes ago, Feathered_IV said:

 they will eventually work their way back around to recognising where the vast majority of their client base is.  

 

At that point there won't be any more planes left to sell to finance the stuff that you want, so they'll close up the shop.

[PatrickAWlson]

12 hours ago, BraveSirRobin said:

 

At that point there won't be any more planes left to sell to finance the stuff that you want, so they'll close up the shop.

 

If you consider a swing to the Pacific there are more than enough planes for the next four or five years.  BoB offers 10 planes easily.  Two land based Pacific modules easily offer another 20.  That gets the team into 2025.

 

One possibility is Great Battles Version 2.  Break backwards compatibility and make a new version of the game that works with any module that the customer has bought, but not necessarily with customers using V1.  Offer better AI and other goodies.  Sell it instead of giving it away.  Would you pay $70 for a significantly better game?  I would in a heartbeat.  Probably no need to go there for several years though.

[PaladinX]

Simply accept that IL2 BOx is mainly an MP flight sim, because the SP is close to unplayable due to missing interaction possibilities with the AI and due to their general behaviour.

Its a very good sim of the aircraft from a technical point of view, of course.

 

But singleplayer im back to IL2 1946 with B.A.T and hundreds of hand written campaigns and missions and a much more believable AI and interaction possibilities...

[LukeFF]

5 hours ago, PaladinX said:

Simply accept that IL2 BOx is mainly an MP flight sim, because the SP is close to unplayable due to missing interaction possibilities with the AI and due to their general behaviour.

 

The numbers (and posts by the developers) don't back up your assertion at all. Plus, when's the last time you flew a singleplayer mission? Yes, nothing's perfect, but career mode missions are quite enjoyable at the moment. 

Edited August 6, 2020 by LukeFF

[PaladinX]

No, they arent enjoyable as long as the AI behaves like it behaves.

Compare to

Sturmo 46

WoFF

F4 BMS

 

Absolutely not competitive in any way to these much older flightsims.

But when i want to enjoy nice graphics, i start BoX for sure.

 

And that makes me sad as there is so much potential.

Edited August 7, 2020 by PaladinX

[PatrickAWlson]

@PaladinX I do not really agree.  Maybe I'm just not as good as you are but I can pretty much only score three victories max, often only one or two.  Sometimes none.  I get shot up often enough.  To me the AI represents an average (i.e. not very good) pilot.  If you want it to represent an ace pilot then yes, it's not there.

 

In combat the issue remains the same as every other combat sim I have ever played.  The AI is just better at turning than handling any form of energy fighting.  There have been nice strides in the right direction but there is still more to do.  

 

The most important thing for me is making combat less lethal.  Pilots simply did not hang around with holes in their planes and smoke pouring from them.  Most did not try to engage when badly out numbered.  Most would try to disengage and escape when they found themselves in a bad situation.  If they were flying a fighter that they knew was faster and heavier they would frequently break off and reset if the initial attack failed, often choosing not to engage again.  Those would go a long way towards making the AI more realistic, but I can imagine the howls of despair when the AI actually does not hang around to be finished (the AI isn't fighting me!),

 

Accuracy: what the AI lacks is ace accuracy.  For a representation of an average pilot it is plenty accurate (the average pilot couldn't hit a moving target).

Flying skill: it is fine for a representation of an average pilot.  Again, it is not ace level.  Getting it to that point is still a WIP.

Thought processes: This is the biggest area of need.  Seemingly little to no concept of team work.  Little to no survival instinct.  Little to no ability to assess a situation beyond the next two seconds.   

 

If 1C perfected the AI - as in made it behave like real WWII pilots - I doubt the average AI would shoot better or fly better.  Higher AI levels certainly would.  What you would have is a lot less combat and much more "I'm out of here".

[AEthelraedUnraed]

On 8/6/2020 at 3:55 PM, PaladinX said:

Simply accept that IL2 BOx is mainly an MP flight sim, because the SP is close to unplayable due to missing interaction possibilities with the AI and due to their general behaviour.

As someone who almost exclusively plays SP, I cannot possibly agree with your point of view.

 

On 8/7/2020 at 8:54 AM, PaladinX said:

Compare to

Sturmo 46

WoFF

F4 BMS

 

Absolutely not competitive in any way to these much older flightsims.

A large part of that can be attributed to the fact that - at least in the case of IL2 1946, not completely sure about the others - the AI uses a different, simplified flight model. Sure, if the AI gets better stall behaviour, more lift and an engine on steroids, it's gonna be more competitive ?

 

On 8/7/2020 at 2:37 PM, PatrickAWlson said:

In combat the issue remains the same as every other combat sim I have ever played.  The AI is just better at turning than handling any form of energy fighting.  There have been nice strides in the right direction but there is still more to do. 

I've found that since the last AI update, energy fighters don't generally try to get into a turn fight with you and will even actively run away from one. Of course this does depend on the AI skill; rookies will generally enter a turn fight no matter what.

 

One of the most frustrating experiences I've ever had in the sim was fighting an ace-level S.E.5 with a Dreidecker. After 15 minutes of dodging his high-speed attacks without obtaining a single firing solution I just gave up and ran away.

 

On 8/7/2020 at 2:37 PM, PatrickAWlson said:

Accuracy: what the AI lacks is ace accuracy.  For a representation of an average pilot it is plenty accurate (the average pilot couldn't hit a moving target).

Flying skill: it is fine for a representation of an average pilot.  Again, it is not ace level.  Getting it to that point is still a WIP.

Thought processes: This is the biggest area of need.  Seemingly little to no concept of team work.  Little to no survival instinct.  Little to no ability to assess a situation beyond the next two seconds.  

Generally agree with you here, although I think flying skill per se isn't the problem - I've yet to see an ace get into a spin for example. I think the problem is too much passiveness; if, for example, you are in a tight turn and are hitting the enemy with your guns, they'll just keep on flying the same turn as before. While for a real pilot it should be clear that the turn isn't working and they should try something else.

 

EDIT: yay, 100th post ?

Edited August 9, 2020 by AEthelraedUnraed

[oc2209]

On 8/5/2020 at 3:16 PM, Yogiflight said:

[1] You can't compare AI behaviour in QMB with its behaviour in career and campaign missions, because it is completely different, what AI does.

 

[2] This was changed some time ago. I can remember very well, how deadly AI fighters were in deflection shooting. They were hitting always in every angle, so you better did everything to not give them a chance to give you a deflection shooting, because then you were dead.

On point 1, I play both Quick and Career. The AI does behave differently, yes; but in my experience, the AI is even weaker versus humans in career because it seems 'distracted' by having other planes in the area, as well as its assigned objectives. In Quick, it has nothing but me to focus on.

 

On point 2, I've simply never seen what you describe, in either Quick or Career modes. Are you saying the AI deflection shooting was better in BoX, but it's been scaled back recently?

 

On 8/7/2020 at 5:37 AM, PatrickAWlson said:

@PaladinX I do not really agree.  Maybe I'm just not as good as you are but I can pretty much only score three victories max, often only one or two.  Sometimes none.  I get shot up often enough.  To me the AI represents an average (i.e. not very good) pilot.  If you want it to represent an ace pilot then yes, it's not there.

 

Accuracy: what the AI lacks is ace accuracy.  For a representation of an average pilot it is plenty accurate (the average pilot couldn't hit a moving target).

Flying skill: it is fine for a representation of an average pilot.  Again, it is not ace level.  Getting it to that point is still a WIP.

Thought processes: This is the biggest area of need.  Seemingly little to no concept of team work.  Little to no survival instinct.  Little to no ability to assess a situation beyond the next two seconds.   

 

If 1C perfected the AI - as in made it behave like real WWII pilots - I doubt the average AI would shoot better or fly better.  Higher AI levels certainly would.  What you would have is a lot less combat and much more "I'm out of here".

I agree with your suggestions and observations here (and how damaged planes would want to disengage immediately). That said, I think the amount of kills we could average on career missions depends on the specific campaign and our plane. I haven't flown much in the early war careers, but the 109G-K with the 30mm is pure murder against AI. When I'm flying in Bodenplatte career, I face Tempests as the sole fighter (barring the occasional Spitfire recon), and they're escorting either Thunderbolts or B-25s. Tempests are tough opponents due to their speed and turning ability, so generally I would be lucky to get 3 in one mission. Thunderbolts, however, are my bread and butter. Since they take virtually no evasive action besides shallow turns, shooting them down from point-blank with 5-10 30mm shots per kill is very easy. I'm positive that between my 30mm and 13mm ammo, I could shoot down 8 Thunderbolts in a career mission. The only thing stopping me is spotting; not the aggressiveness of the Tempest escorts. That said, I've shot down 6 Tbs. in one mission. If the Tempests attacked me with any kind of regularity, I'd obviously have a much harder time with the Tbs. That leads me to my next points.

 

Per your mention of AI accuracy, I think AI pilots need 2 layers of skill. They need inherent skill, and acquired skill. Approximately 1/4 of pilots should have higher than average accuracy and aggression (I'm not saying that fraction correlates to reality; but it'd be a way of ensuring you'd meet at least one good pilot per combat); this inherent skill should be augmented by experience, so that when natural skill and experience coalesce, a 'super ace' AI pilot would emerge. This would apply to your wingmen too.

 

So, for example, a naturally skilled pilot would always start at regular quality; never rookie. At veteran level of experience, they'd fly like an ace. And at ace, they'd be better than aces who only became so by experience. Maybe a special designation, like 'Ace+' would be in order. I know that sounds pointlessly complicated, but it does reflect reality; just as all pilots weren't equal, all aces weren't equal. There's a huge difference between an ace who got 5 kills in 50 combats, where several of those could be bombers or obsolete planes; and an ace who got 50 kills in 150 combats, most against fighters.

 

Something else the AI needs to do at lower skill levels: spray fire in your direction. This would be very effective in chasing you off of escorted bombers and ground attack planes. An AI with the inherent skill trait would be more accurate at any range (and extremely accurate up close), while regular AI would be wildly inaccurate and wasteful, firing from extreme ranges until reaching veteran or higher levels of experience. At ace level, the AI wouldn't fire until it was sure it would hit you; meaning just like in real life, you wouldn't have the luxury of seeing tracers going by your wing. You'd just be flying along obliviously, then suddenly you'd hear metal thuds and everything would go dark.

[PatrickAWlson]

@oc2209 Expected victory count brings into question what one views as realistic.  The best pilots generally managed a victory every 2 combats.  Here we're talking Erich Hartmann and Robert Johnson.  Very good pilots might get a victory  every 10 combats.  These pilots were aces.  An American with 100 missions and 50 combats might end the war with 5 kills - not bad.  A German with 1000 missions and 500 combats might end the war with 50 - also not bad.  Most pilots scored zero victories.  Scoring three per mission on average and sometimes as many as 6 makes you pretty much the best pilot of the war by a long shot.

 

The means for pilot progression already exist and both PWCG and in the game career mode use it.  Pilots start as novices and then improve as they gain missions and victories.  Eventually they top out as aces.  I do agree that these needs to be another layer above ace - expert - whatever.  For me I would generally reserve that for historical aces.  However, the current crop of AI flies and shoots mostly like normal pilots.  I would argue that the current AI really does not often l fly like an ace.

 

Also agree about spraying fire.  One mark of the newbie was to start shooting at 600 yards and miss badly.  

 

I am still on about sense of survival as being the #1 thing that I would like to see.  These planes are not that easy to down.  If AI pilots bugged out in a realistic manner kill counts would go way down,  You would either have to chase your cripple to finish him or let him go.

Edited August 9, 2020 by PatrickAWlson

[oc2209]

5 minutes ago, PatrickAWlson said:

@oc2209 Expected victory count brings into question what one views as realistic.  The best pilots generally managed a victory every 2 combats.  Here we're talking Erich Hartmann and Robert Johnson.  Very good pilots might get a victory  every 10 missions.  These aces were aces.  An American with 100 missions and 50 combats might end the war with 5 kills - not bad.  A German with 1000 missions and 500 combats might end the war with 50 - also not bad.  Most pilots scored zero victories.  Scoring three on average and sometimes as many as 6 makes you pretty much the best pilot of the war by a long shot.

The number of Germans who actually survived the war with 500 combats is quite small I imagine. I wager a lot more of the less famous experten got their 50-100 kills in fewer sorties and were dead by the end of '43. In other words, they burned out quickly. What makes Hartmann exceptional is that he got the bulk of his kills when the Germans had lost most of their early war advantages.

 

What I find interesting about the accounts of pilots from all nations with exceptionally high kill counts in one sortie or one day: they rarely went on to have exceptionally high scores overall. I think most of those cases are a matter of profound luck meeting higher-than-average skill. But luck gets the lion's share in my opinion. Getting consistent kills requires preternatural awareness, better than 20/20 vision, aggression, the gunnery skills of a lifelong hunter, physical strength and endurance, and luck besides all that.

 

I have no illusions about my pilot/gunnery skill in career mode applying to real life, or even applying to multiplayer. The only thing I do that many real aces did: I fire from point-blank range. Hartmann and others have noted that some aces can fire from exceptional ranges and consistently score kills; but they are even rarer than 'average' aces who use the point-blank technique. The only thing wrong with 30mm kills at close range, is the pieces of tail that invariably fly in your face. I would estimate 80% of my P-47 kills result from a sawed-off vertical stabilizer. The pilot's often killed at the same time, rendering the structural damage and subsequent spin redundant.

[Yogiflight]

1 hour ago, oc2209 said:

Are you saying the AI deflection shooting was better in BoX, but it's been scaled back recently?

It seems to be that way, however not too recently. It is quite some time, that I didn't see that anymore. There was a lot of complaining about that superhuman deflection shooting, also by me. There is nothing wrong with good deflection shooting by experienced AI pilots, as long as it is reasonable. But this might be hard to program.

[oc2209]

1 hour ago, Yogiflight said:

It seems to be that way, however not too recently. It is quite some time, that I didn't see that anymore. There was a lot of complaining about that superhuman deflection shooting, also by me. There is nothing wrong with good deflection shooting by experienced AI pilots, as long as it is reasonable. But this might be hard to program.

Hmm, interesting.

 

I've only been playing BoX for a few months now, so I had no idea AI was ever capable of deflection. Evidently, what we're seeing now is a bit of an over-correction, perhaps. Maybe once other AI routines are established or refined, they can reintroduce a modified deflection behavior.

[oc2209]

On 8/9/2020 at 12:27 PM, PatrickAWlson said:

I am still on about sense of survival as being the #1 thing that I would like to see.  These planes are not that easy to down.  If AI pilots bugged out in a realistic manner kill counts would go way down,  You would either have to chase your cripple to finish him or let him go.

 

Okay, since our conversation sparked my curiosity, I just returned to my career after leaving it for most of the last month. I had 128 kills in 45 sorties, in Bodenplatte. This is regular career mode, not your mod, by the way.

 

An important note about my kills per sortie rate is that I made myself commanding officer from the beginning, so I choose which missions I fly or not. I totally avoid anything that isn't a free hunt or interception. So that right there will automatically make my rate artificially high.

 

In terms of AI trying to escape, that might work with early 109 armament or any planes with wing mounted guns. But when a nose mounted 30mm is fired from under 100 yards, there is no 'crippled' condition; it's either miss or instantly fatal. They all strike from engine to tail, depending on deflection. I rarely waste 30mm on the wings.  

Edited August 11, 2020 by oc2209
Too long and boring.

[oc2209]

While my detailed combat reports are only interesting to me, I am actually trying to come up with a viable strategy to keep wingmen alive, and secondarily to study AI behavior.

 

In my next experiment, I took 1 wingman on an intercept mission. We faced 6 Thunderbolts and 8 Tempests. I ordered him to follow me (as opposed to cover me). First attempt, I shot down two planes, but the third, a TB, broke up in front of me and ripped my wing tip off. Restarted. Second attempt: Shot down 1 TB, was chasing a Tempest, shooting poorly, failed to kill him in one pass, and as I was slowed up to match his jinking, another Tempest blew my wing off from directly behind (I was flying slow and level, pretty much; any kind of turning makes you invincible to them).

 

Third attempt: I now order my only wingman to cover me, just to see if there's any difference in his behavior. No discernible difference. This time, the battle goes well for me regardless. There are two Fw-190D in the area by happenstance, which provides a great distraction (seemingly better than the first two times). I keep my wingman alive, including shooting a Tempest off his tail as he attempts a shallow, high speed climb to escape said Tempest (sigh). He's pretty damaged afterwards, but still looks flyable. I order him home 3 times, no response. Next time I check on him, he's vanished. As I later found out, he crashed fatally. Note: he was a complete rookie with zero combat experience. Maybe he'd have done better in this scenario with experience. 

 

Next mission, I go completely solo on a bomber intercept (A-20s). Oddly, these bombers have no escort at all. A second solo intercept also has no escorts. Career difficulty is set to 'moderate,' by the way. I have no explanation for the lack of escorts in these missions.

 

Finally, I get a free hunt. I go solo. I meet 6 Tempests. I quickly shoot one down, and then lure the remaining over German territory (AI will not shoot at you at all, if you fly low enough; regardless of your speed). I then enter a perpetual loop on the deck, with the goal of letting my emergency power recharge. As I'm in said loop, in quick succession, 4 Tempests collide (in 2 pairs, independent of each other) as they are looping around me. That leaves 1, which I then shoot down with emergency power.

 

If anyone's bothered to read this much (I don't blame anyone who didn't), this is my conclusion:

 

If you want to lose fewer wingmen, take fewer wingmen in your flight. Going solo isn't practical, as you can't rely on most of an enemy flight colliding. So I'd advise taking 1-3 wingmen at most. Simply put, I can't think of a single way to keep them alive, other than taking fewer into combat to minimize catastrophic losses. Taking fewer along also does theoretically make it easier to keep track of them. 

 

Beyond that, I suggest starting a career as commander, and then sending full AI flights on any missions you don't want to fly (like escort and ground attack, for example). I've noticed that AI are rarely killed (but often wounded or get damaged planes) when you send them on a mission without you. This could be a way of giving them experience without endangering their lives. Then, in missions where you do fly, take fewer wingmen; but maybe take only your most experienced pilots to compensate.

Edited August 11, 2020 by oc2209

[PatrickAWlson]

On 8/9/2020 at 4:24 PM, oc2209 said:

The number of Germans who actually survived the war with 500 combats is quite small I imagine ...

 

Not really.  As I look through Germany's top aces it seems about half survived.  This wiki link shows the aces by victories and shows who survived and who did not.

 

https://en.wikipedia.org/wiki/List_of_World_War_II_flying_aces

 

My impression is that the western front was far more deadly than the east.  

Edited August 11, 2020 by PatrickAWlson

[LukeFF]

1 hour ago, PatrickAWlson said:

My impression is that the western front was far more deadly than the east.  

 

Quite so. I'm reading through volume 2 of JG 26 War Diary right now, and it's noted that for the month of June 1944 on the Normandy front, the Germans lost 3 fighters for every 1 Allied fighter shot down, and loss rates per mission were running at 20-30%. It was a brutal place to be a German pilot.

Edited August 11, 2020 by LukeFF

[oc2209]

6 hours ago, PatrickAWlson said:

 

Not really.  As I look through Germany's top aces it seems about half survived.

We simply have different definitions of 'quite small,' then.

 

To make a conclusive list, we'd have to find the total sortie count, the combat sortie count, and kills, per ace. The probability of combat is pretty simple; the more you fly, the more you die (this is obviously truer for the losers in a war more than the victors). Exceptions to this rule are statistical anomalies; whether they number in the single digits or several dozen makes little difference in my mind.

 

My assertion is, and was, that most German pilots with around 50 kills (going by the hypothetical number you originally mentioned) did not take 1000 sorties or 500 combats to achieve that kill count. The basis for my argument is that few pilots would live that long. Thus, probability would be in favor of them scoring quickly and dying quickly.

 

6 hours ago, PatrickAWlson said:

My impression is that the western front was far more deadly than the east.  

 

Does anyone have statistics of total German pilot death count in the east versus the west? I've never seen a clear number, so I'm curious.

 

Just because a relatively small number of German aces amassed huge scores on the Eastern front, doesn't necessarily mean it wasn't perilous to average German pilots.

 

Going back to the OP, I ran another test.

 

I put a 1941 career (in a 109F-4) on hard (to see if that made any difference in friendly AI behavior), then took an ace AI as my only wingman. We met 2 MiG-3s. Incidentally, I always remove my extra back armor because it blocks the rear view too much; we would've been bounced had I not removed the armor.

 

I shot down a MiG. My wingman says 'target destroyed' over the radio. A few seconds later, he's vanished, along with the second MiG. In the post-mission brief, the map showing where he was downed doesn't say he was shot down. I can only conclude he collided with the remaining MiG, as implausible as that seems to me.

 

So, once again, I can't think of a single way to keep friendly AI alive. When the top ace in my roster dies inexplicably, for apparently no good reason, in a 2v2 combat... yeah.

 

Another fun anecdote from this mission: as I was landing back at my base, I was strafed by 2 I-16s that I hadn't noticed were following me. If they'd hit me as I was landing, without my back armor, I'd have been dead for sure (the attacks were from behind). Luckily they missed me. I took off again (ramming the throttle to emergency almost pulled me off the runway) to engage them, only to find that 3 MiGs had also arrived on the scene. No friendly planes, and the AA guns were awful shots. Hard mode indeed.

 

The AI might be dumb in a few areas, but it was one hell of a fun sortie anyway. 40 minutes long.

Edited August 12, 2020 by oc2209
Didn't intend to merge my posts. Oh well.

[Tinku74]

I imagine that machine learning could provide a pathway to realistic AI. 
 

Have the AI observe a huge number of online PvP dogfights, and “learn” how players act and react in various situations, with different combinations of adversary planes etc. 
 

You will have a normal distribution of skills, a few Über pilots, a few complete newbies, and a vast majority of average pilots. 

But I assume that if this were  is viable, it would already have been attempted. 
 

I for one welcome and appreciate that the team took up the improvement of AI, even if it is in small, incremental steps. 

 

[PatrickAWlson]

36 minutes ago, Tinku74 said:

I imagine that machine learning could provide a pathway to realistic AI. 
 

Have the AI observe a huge number of online PvP dogfights, and “learn” how players act and react in various situations, with different combinations of adversary planes etc. 
 

You will have a normal distribution of skills, a few Über pilots, a few complete newbies, and a vast majority of average pilots. 

But I assume that if this were  is viable, it would already have been attempted. 
 

I for one welcome and appreciate that the team took up the improvement of AI, even if it is in small, incremental steps. 

 

I don't think it's really a machine learning thing, at least not at the PC game level.  AI and combat techniques of WWII are known.  There really is no discovery.  

 

Much more of a state machine with  transitions, with some transitions limited by configured AI capabilities (novice - ace).  Think Turing.  Throw in some randomization so the AI does not always do the same thing and then you have something.  This is really complex.  Each state (attack air, attack ground, evade air to air, evade ground, break away, fly formation, cover, land, etc) can be made of smaller states and finally decisions.  Once the AI commits to an action it should follow through.  

 

Example: in air to air mode the AI would first decide to attack, evade or set up.  Once in attack, evade or set up it would then decide how to do that.  That might get it down to a specific maneuver that the AI believes will accomplish its goal.  The AI would then perform that maneuver.  All paths would be open to ace, with some paths pruned for lesser AI.  No real learning to do, but a very complex decision tree.

 

Example: Evasion might transition to aggressive vs safe evasion.  The first being an attempt to turn the tables, the second to get out of Dodge.  An aggressive evasion would involve something that gains altitude to flip the situation and hopefully allows transition to attack.  A safe evasion would probably transition to breaking off combat.  A novice might only have two options: turn or dive away.  An ace might do things like half loops, split-S, wingovers and yo-yos.  

 

If there was agreement on the method i think it could be fleshed out over time.  Start by implementing the basic state structure and then focusing on one - let's say attack.  Implement that single state (which is IMHO more or less what we have today).  Make break away what we have today: damaged RTB.  Make evade what we have today - some random maneuver.  Make setup what we have today - limited.  But now you have states, even if they are not well formed.  Over time improve each state and the transitions between states.

 

Example: today if you shoot at a landing plane it will continue to land.  This would be taken care of in the future by allowing the AI to transition from landing state to evade or aggressive.

 

Lots to be done but IMHO the most important thing is deciding on a AI framework and injecting the current AI into that framework.  Then proceed.

[PatrickAWlson]

So I went to the mountains and with three of my buddies ...

 

and the all crashed into the side of a hill while I was merrily strafing a train ...  got to admit the FW190 AI is getting me down.  

[flagdjmetcher]

11 hours ago, PatrickAWlson said:

I don't think it's really a machine learning thing, at least not at the PC game level.  AI and combat techniques of WWII are known.  There really is no discovery.  

 

Much more of a state machine with  transitions, with some transitions limited by configured AI capabilities (novice - ace).  Think Turing.  Throw in some randomization so the AI does not always do the same thing and then you have something.  This is really complex.  Each state (attack air, attack ground, evade air to air, evade ground, break away, fly formation, cover, land, etc) can be made of smaller states and finally decisions.  Once the AI commits to an action it should follow through.  

 

 

I think ML could be useful.  A state machine is fine, but knowing what state you're in is the hard part.  A human can look at a flight of allies engaging a flight of enemies, and have a reasonable idea of who is engaging who and prioritize enemies based on level of threat to self and wingmates.  That's really, really hard to do via a state machine, but quite a reasonable problem for ML.  Once you've sorted out what's happening and where you can best intervene, I agree that the next steps can be a bit more deterministic.

[AEthelraedUnraed]

16 hours ago, Tinku74 said:

I imagine that machine learning could provide a pathway to realistic AI. 
 

Have the AI observe a huge number of online PvP dogfights, and “learn” how players act and react in various situations, with different combinations of adversary planes etc. 
 

You will have a normal distribution of skills, a few Über pilots, a few complete newbies, and a vast majority of average pilots.

As a postgrad in a very machine learning-heavy field of engineering, and with extensive knowledge of different methods of machine learning (including neural networks, the current "hype") and their properties, I don't think machine learning is the best solution here. Sure, it would *eventually* lead to proper behaviour. However:

 

- You have to design it. You have to come up with a way to quantify things such as the position, speed and direction of nearby allied and enemy aircraft, whether they are firing and if so their target, the amount of aircraft, the location on the map, the progression of this all through time, etc. It's likely there's not currently a suitable architecture, meaning that you'd have to design it from the ground up.

 

- You have to train it. This will require hours and hours of recordings. Also, you'd basically want to train an AI to make correct decisions in response to a changing situation (enemy spotted, enemy starts firing at you etc.) while in fact those decisions are rarely made in real life. Mostly, you just stay on the tail of your enemy or execute the Immelmann that you decided to make a couple of seconds ago. You can either have some human decide at which points decisions are actually being made, with the resulting action taken, or run the risk that you're actually training your AI to just keep doing whatever it is doing, since that is usually the best course of action.

 

- It's slow. Take for instance GoogLeNet, one of the popular neural network architectures these days (just as an example; I'm not suggesting it would be particularly suited to a gaming AI). Running data through the network requires some 1,500,000,000 operations (mostly multiplications and additions) per their paper. Including stuff such as retrieving the data from memory, this translates to many times that amount in CPU/GPU operations. Fine if you want to play a game of chess, not so fine for a flight sim with possibly tens of planes that are all making these decisions.

 

3 hours ago, flagdjmetcher said:

I think ML could be useful.  A state machine is fine, but knowing what state you're in is the hard part.  A human can look at a flight of allies engaging a flight of enemies, and have a reasonable idea of who is engaging who and prioritize enemies based on level of threat to self and wingmates.  That's really, really hard to do via a state machine, but quite a reasonable problem for ML.  Once you've sorted out what's happening and where you can best intervene, I agree that the next steps can be a bit more deterministic.

Things such as machine-learning based threat detection are actually still state-of-the-art (a nice example (June 2018) is https://www.youtube.com/watch?time_continue=3&v=zYypJPJipYc). There's research groups, both at universities and in business, that work basically full-time on similar things, so I think it's a bit far-fetched to expect from the Devs that they design such a system in addition to all the other work they do on the sim ?

 

Besides, I don't quite agree with your statement that it's extraordinarily hard to prioritise based on threat level in a state machine. It can be as simple as just adding threat levels based on some of the following properties:

- speed/height (=energy) relative to you

- position relative to you

- aircraft type

- is it firing?

- heading relative to you

 

It's not too hard to program your AI in such a way that a high, fast target flying away gets less attention than a bandit on your six with blazing guns.

[PatrickAWlson]

9 hours ago, flagdjmetcher said:

 

I think ML could be useful.  A state machine is fine, but knowing what state you're in is the hard part.  A human can look at a flight of allies engaging a flight of enemies, and have a reasonable idea of who is engaging who and prioritize enemies based on level of threat to self and wingmates.  That's really, really hard to do via a state machine, but quite a reasonable problem for ML.  Once you've sorted out what's happening and where you can best intervene, I agree that the next steps can be a bit more deterministic.

 

The AI already runs on a state machine of sorts.  It is simply IMHO not a terribly well fleshed out one and I suspect that the state transitions are not well defined or often entirely missing.  My proposal is really an incremental improvement on what they have and not a scrap it and start over.  

 

ML is all about thousands/millions/billions of repeats to refine the best course of action - thus the "learning" part.  The IBM chess program spend years learning before it was actually let loose to play.  Similar to the chess program, I guess an ML system could be trained in house and then turned loose.  I assume that is where you would go with this.  That sounds a whole lot harder than a state machine but I must admit that I may be seeing the problem as a nail because I know how to use a hammer.  I am very comfortable with complex state transitions.  I admit that I lack expertise in ML.

 

I also must admit that I have a failure bias towards ML.  I have worked at several organizations, including one very well known one, where there has been all sorts of talk about ML and very little action.  I have yet to see or hear of a small organization pull off a major feat like successfully teaching a computer WWII combat techniques.  Most "ML" that I have seen has in fact been statistics and randomization, not true ML.

 

Edited August 13, 2020 by PatrickAWlson