AI planes

[1PL-Husar-1Esk]

10 minutes ago, Aapje said:

Both Il-2 and DCS are already products designed primarily for multi-player, which don't provide an optimal single player experience

Wrong , il2 primary is a single player game.

I'm for AI trained from players gameplay.

Edited February 29, 2024 by 1PL-Husar-1Esk

[CzechTexan]

I have also become more aware of unusual A/I behavior.  I do like the more aggressive head-ons.  However, the deadly sniper shots seem to be more frequent.  There are not even any rat-tat-tat hits on the aircraft to warn the pilot - just a single shot to the cockpit!  It's very frustrating.

[357th_KW]

I've noticed the AI fighters are definitely hitting a lot more shots.  I'm not sure this is such a negative thing - in the past I've had to use AI's set to Ace or High almost exclusively in mission building, as any other setting was just a helpless target.  And even on those settings the AIs really weren't much of a threat to anyone used to playing against humans.  You could generally just focus on one at a time, shooting down a whole formation with very little danger from the others.  Now I'm finding you definitely have to pay more attention to them, and treat them as a potential threat like you would playing against humans. 

[Dagwoodyt]

I'm learning the P47D's flying 1v1's vs 190A3 or 110G-2's. I keep AI set to "Ace". I am comfortable that the AI aim accuracy is reasonable. The AI will make me pay if I get sloppy so I have to be conscientious about taking evasive action early even at cost of position.

[Lusekofte]

2 hours ago, CzechTexan said:

I have also become more aware of unusual A/I behavior.  I do like the more aggressive head-ons.  However, the deadly sniper shots seem to be more frequent.  There are not even any rat-tat-tat hits on the aircraft to warn the pilot - just a single shot to the cockpit!  It's very frustrating.

It is odd how they always pk you from near impossible angles without any hits other places in the plane every time. But when they are on the six you get a lot of chances. That bit is way off and not comprehendible. But they make good opponents lot of the time at least until they run away after a while. 
I fly underdogs and can not easily follow them when they stop being aggressive 

[SYN_Speck]

9 hours ago, 1PL-Husar-1Esk said:

Wrong , il2 primary is a single player game.

I'm for AI trained from players gameplay.

I think you're correct in that most users of combat flight sims are single players.

 

But, I think Aapje is correct that "Il-2 [is] already [a] product designed primarily for multi-player" (I don't know about DCS).

 

Having participated in ROF from the start (and played IL2 since 2005) and having the rare and really fun experience of being involved in a fairly small community where the devs shared a lot with us, I know for sure that ROF was designed for online play first, with SP an afterthought. And it's my understanding that the current IL2 series evolved from ROF. It makes sense that devs from a small independent studio, as those who developed ROF, would build a product that first and foremost allowed them to have a fun time flying with and vs. their friends in MP. The SP experience in ROF and subsequently in the GB series has never worked all that well, and while the details of squadron and aircraft assignments and attempts at immersion through pilot stories etc have improved over the years, the AI just hasn't. it's a shame. So many elements of the flight sim experience in general, too, are leaps and bounds better than they were in 2009 (release of ROF) or earlier in IL2 FB. The FMs, the graphics, etc are are improved by generations. But the AI is, if anything, WORSE than it was in 2005 when I started! It's so frustrating.

 

I really think that the sim must attract lots of interested people, and a few of them who have the time and devotion to have a satisfying experience online may stick around, but I just have to think that so many SP players soon drift off to other sims or computer games.. because the AI is SO annoyingly unrealistic! All the devs ever seem to do is fine-tune how accurately the AI shoot, and from how far away.

 

For the sim to thrive in future the AI just has to improve significantly. I really hope the devs can manage to figure out how to do it. I think Aapje has a lot of good ideas butI know it's way easier said than done.

[1PL-Husar-1Esk]

12 hours ago, SYN_Speck said:

I think you're correct in that most users of combat flight sims are single players.

 

But, I think Aapje is correct that "Il-2 [is] already [a] product designed primarily for multi-player" (I don't know about DCS).

 

Having participated in ROF from the start (and played IL2 since 2005) and having the rare and really fun experience of being involved in a fairly small community where the devs shared a lot with us, I know for sure that ROF was designed for online play first, with SP an afterthought. And it's my understanding that the current IL2 series evolved from ROF. It makes sense that devs from a small independent studio, as those who developed ROF, would build a product that first and foremost allowed them to have a fun time flying with and vs. their friends in MP. The SP experience in ROF and subsequently in the GB series has never worked all that well, and while the details of squadron and aircraft assignments and attempts at immersion through pilot stories etc have improved over the years, the AI just hasn't. it's a shame. So many elements of the flight sim experience in general, too, are leaps and bounds better than they were in 2009 (release of ROF) or earlier in IL2 FB. The FMs, the graphics, etc are are improved by generations. But the AI is, if anything, WORSE than it was in 2005 when I started! It's so frustrating.

 

I really think that the sim must attract lots of interested people, and a few of them who have the time and devotion to have a satisfying experience online may stick around, but I just have to think that so many SP players soon drift off to other sims or computer games.. because the AI is SO annoyingly unrealistic! All the devs ever seem to do is fine-tune how accurately the AI shoot, and from how far away.

 

For the sim to thrive in future the AI just has to improve significantly. I really hope the devs can manage to figure out how to do it. I think Aapje has a lot of good ideas butI know it's way easier said than done.

I'm aware that both games started with multimeter mode only, this is a business model. You can build, test and sell product faster. You start generating revenue faster before you finish product as a whole game. Single player content is  time consuming. Why it is single player game now - becouse ~80 % of buyers do play single player mode only. This information is from devs. So if a game generate more revenue from people playing single player, you can't say this game it is a product design primary for multiplayer or that happened by chance not by devs intent. I believe they have that route planed.

[Gambit21]

12 hours ago, SYN_Speck said:

I think you're correct in that most users of combat flight sims are single players.

 

But, I think Aapje is correct that "Il-2 [is] already [a] product designed primarily for multi-player" (I don't know about DCS).

 

 

Most users are single player. However from a development/decision standpoint you’re more correct about this than you know. The state of the sim in more than one area, for years now, reflects this as well…AI and career logic included.

 

I’m not calling this wrong mind you, however I’m hoping to see things improve in the upcoming iteration.

 

 

 

 

 

 

[Jaegermeister]

14 hours ago, SYN_Speck said:

The SP experience in ROF and subsequently in the GB series has never worked all that well, and while the details of squadron and aircraft assignments and attempts at immersion through pilot stories etc have improved over the years, the AI just hasn't. it's a shame. So many elements of the flight sim experience in general, too, are leaps and bounds better than they were in 2009 (release of ROF) or earlier in IL2 FB. The FMs, the graphics, etc are are improved by generations. But the AI is, if anything, WORSE than it was in 2005 when I started! It's so frustrating.

 

I really think that the sim must attract lots of interested people, and a few of them who have the time and devotion to have a satisfying experience online may stick around, but I just have to think that so many SP players soon drift off to other sims or computer games.. because the AI is SO annoyingly unrealistic! All the devs ever seem to do is fine-tune how accurately the AI shoot, and from how far away.

 

For the sim to thrive in future the AI just has to improve significantly. I really hope the devs can manage to figure out how to do it. 

 

The biggest hurdle to creating a realistic and immersive single player experience is the radio comms interface, formation commands, AI formation flying and something even close to real world sword and shield tactics, which is what every Air Force eventually used in WWII. We still don't even have flights divided into elements to cover each other.

 

Damage models, ballistics and gunnery have been repeatedly adjusted while the other issues have never been addressed at all since adding British and American voices with Battle of Bodenplatte. It is a big missed opportunity in my opinion

[Aapje]

I think that the first sim to implement a machine learning AI, rather than a procedural AI, will have a killer feature that will draw a lot of players, even if the AI is 'only' much better than procedural AI and cannot do the truly advanced things.

 

My advice would be for the sim companies to start experimenting with it soon, since figuring out what data you need and how you can optimize the models is going to be a learning process that is probably going to take way more waiting time than developer time, as you need to wait for the data to come in from multiplayer servers/gamers and for the models to get trained, so starting early can then be cheaper than having to catch up in an inefficient way. After all, while waiting for those things, the developer can just do something else. If you need to catch up, the developer may need to start new experiments before the learnings from previous experiments are even in, causing him to waste time on experiments that were never going to work.

 

I don't do AI models, but my understanding is that the normal process is to basically experiment with creating relatively small and lightly trained models on a single GPU, like a 4090 or 4080, and only once you create a good small model, train the big production model on multi-GPU hardware at higher expense. Although the costs should be a fraction of the big models like ChatGPT, since the scope is way smaller.

 

That first step just using a local machine shouldn't be too expensive, aside from the developer time, but it should be a very good learning experience to figure out what seems to me to be the inevitable future of simming.

[Parazaine]

I have only just got into IL2, although i played RoF for a few years and the worrying thing is the exact same single player issues that plagued RoF (and might still do for all i know) seem to plague IL2 also.

 

I think i read somewhere that the AI was, in some part an iteration of RoF's? If so then its not likely to get fixed any time soon. For all, those saying IL2 has a mostly SP player base, that may well be true but there was always an elitist feel to the developers in RoF before Jason joined the team and it seemed fairly obvious that their focus was on multiplayer. True or not, from a, until recently, outside view, it seems the same for IL2 in general.

 

If true, then changes in 'bugs' in SP that have been around for years are probably unlikely to be a focus...sad if, as others have stated, the majority of the player base are SP. My own experience in the last week or so has been a frustrating one in SP career or the excellent PWCG mod...i have been sniped by one kill shot from rear gunners, AAA and unknown sources in what seems an unbelievable way

 

I have also read the comments from people who say the AI is too easy, or they have no problems. This is unhelpful, just look at all the calls for change from other players....some middle ground solution must be possible....a difficulty slider / accuracy slider. Admittedly, i am not a coder so i dont know how hard or easy this is to implement....i only know that i have tried all the mods that purport to help and they dont seem to make much difference currently

[RedeyeStorm]

You guys are making far too many assumptions. The problems with the AI have been acknowledged for a long time. The problem with solving it is not a lack of willingness but requires a dedicated AI programmer which to company has been unable to find and retain. This has been communicated on this forum for several years. 

[CountZero]

considering they are making new game i dont expect they gona be doing any big meaningfull AI improvments here, if its so hard to fined AI specialist why would they waist his time on this game insted making new game AI good from scrach.

 

 

cooperative campaign should be their focus also, you able to play SP missions, campaigns careers with frends vs AI, for that you have to have belivable AI and control over it. Just see how popular this new shooter is today, and its coop MP, not ppl vs ppl its ppl cooperating vs AI, you cut all the toxisity that drives SP players off MP and ppl have fun, you dont have to spend 6+h a day playing just so you can be good in MP, when you can have coop with guys on your level, and its mutch more interesting playing SP with 2 or more frends then alone with bad AI, you see flaws in AI cooperation then more.

 

[giftgruen]

This is an interesting discussion.

 

First of all: Yes. LLM is the way to go on long term to simulate intelligence in a way that we could call 'acceptable simulation of human intelligence'

However - a human pilots behaviour is not only controlled by knowledge and decision finding, a huan pilots behaviour is also ( largely ! ) influenced

by factors like mood, fear, braveness, feeling tired, beeing unconcentrated, beeing in anger, beeing cool ... and so on. Emotions. Character. 

 

So even if LLM could enable us to create bots with perfect intelligent behaviour from a 'technical abilities' point of view, this is not the

same like creating bots that simulate human pilot behaviour.

 

In my opinion, actual LLM ( though achieving a real breakthrough in AI ) is still far away from what we dream of.

Let alone having a LLM generated human behaviour simultanously for multiple bots, limited by a home computers CPU.

 

Yes, for sure, some day in future these limitations wont exist any more.  

But for now and for the next years - we are probably stuck with procedural AI, at least when talking about AI in simple computer games.

 

Discussing AI in our actual context IL2 therefore means, that we need to discuss procedural AI.

Is it possible to create a better procedural AI than IL2 does at the moment ?

I am absolutely convinced that the answer is YES.

 

IL2 is far off the limits of procedural AI.

 

Yes. It would be nice to have (LLM) bot pilots with a perfect simulation of HUMAN behaviour.

But for the moment I'd be happy to have bot pilots that have acceptable BOT behaviour

[1PL-Husar-1Esk]

Guys, you can find answers why it's difficult for programers to train AI. BTW  LLM stands for language models, good for RPG AI NPC dialogue maybe but not for combat flight sim.

 

 

Edited March 2, 2024 by 1PL-Husar-1Esk

[Parazaine]

I don't think AI 'learning' is the answer...we are not trying to make the AI more efficient,only more human and therefore less efficient

 

Random number generators have been around long enough to be used to simulate human inaccuracy surely? Human-like behavior is another matter

 

I remember seeing a US Air force Survey about B17 gunnery accuracy and it was surprisingly low in the single figures on average. Also a study of casualties and wounds n relation to ordinance fired reveals much the same thing from all combats. I will try and find some evidence to back up these points

 

Making the AI behave in a more human way is obviously far more difficult

Edited March 2, 2024 by Parazaine

[1PL-Husar-1Esk]

5 minutes ago, Parazaine said:

I don't think AI 'learning' is the answer...we are not trying to make the AI more efficient,only more human and therefore less efficient

Depends from which player group AI would copy behavior, you can learn from less efficient players and have AI less efficient.

[Aapje]

5 hours ago, RedeyeStorm said:

The problem with solving it is not a lack of willingness but requires a dedicated AI programmer which to company has been unable to find and retain. This has been communicated on this forum for several years. 

If they are going to use machine learning, there's probably going to be a huge shortage of experienced developers for quite a while, so a more realistic option is to let someone learn this stuff on the job.

[Aapje]

2 hours ago, 1PL-Husar-1Esk said:

Guys, you can find answers why it's difficult for programers to train AI. BTW  LLM stands for language models, good for RPG AI NPC dialogue maybe but not for combat flight sim.

You don't actually need a full LLM to have a machine learning AI that can fly similar to people. You only need something like that if you want the AI to be able to understand and/or use real language.

 

But you should be able to build an AI that can fly, shoot, etc fairly realistically, can follow orders and can at least interpret and send some standardized messages, with a relatively small and relatively easy to run model. The AI model I'm thinking off would perhaps have something like 10's of thousands of parameters, while something like ChatGPT 4 has 1 trillion parameters. That's why ChatGPT costs many millions to train and is very demanding to run as well, completely unfeasible for a small company like 1GCS.

 

So a relatively small model is the direction I would go in. With programming, it's typically true that 90% of the desired features take 10% of the time, and the other 10% of features take 90% of the time. So if you can get away with it, the cost-effective option is to only make the things that are relatively easy. I think that is the only realistic option in this case and would result in a solution that is of course not perfect, but much better than the existing AI.

2 hours ago, Parazaine said:

I don't think AI 'learning' is the answer...we are not trying to make the AI more efficient,only more human and therefore less efficient

 

If you look at my proposal, then the idea is to train the model to mimic humans, not to train it to be the best pilot (although it could of course still mimic the better human pilots).

 

My proposal also includes skill level of the human pilots as a parameter for the model, which means that the model can learn how human pilots of different skill levels fly and you can then ask it to fly at a certain skill level.

 

2 hours ago, Parazaine said:

Random number generators have been around long enough to be used to simulate human inaccuracy surely? Human-like behavior is another matter

 

Human inaccuracy is not random and like you say, inaccuracy is only one aspect of human behavior..

[Aapje]

3 hours ago, giftgruen said:

However - a human pilots behaviour is not only controlled by knowledge and decision finding, a huan pilots behaviour is also ( largely ! ) influenced

by factors like mood, fear, braveness, feeling tired, beeing unconcentrated, beeing in anger, beeing cool ... and so on. Emotions. Character. 

If the parameters to the model include information that results in things like anger or fear, then the model can reproduce these things, at least to some extent.

 

For example, I think that it is likely that human pilots who get damaged are more likely to want to take revenge, or want to flee (fight or flight response). So if the damage level of the plane (and how long ago they were damaged) are used as parameters of the model, it can learn how human pilots react to being damaged.

 

The big benefit of a machine learning model over procedural AI is that the the former will automatically figure out most of these behaviors, once you feed it with the right data and enough data, while a procedural AI requires all of these behaviors to be hand-coded and hand-tuned.

Edited March 2, 2024 by Aapje

[Letka_13/Arrow_]

On the other hand, AI for offline flying in career should not represent some challenging combat bots that model online PvP behavior, which is more about an e-sport mentality.  I doubt that any sane WW2 pilot would try to get revenge after being hit, but rather bug out and try to make it back. 

Good AI in a historical combat sim is in my opinion more about using realistic tactics, which results in immersive experience. And historically it was more about survival, good positioning, using your tactical advantage to the maximum, short but intensive dogfights and generally realistic attrition rate with realistic chances of making it through by flying smartly. This is in my opinion not that much about exceptional flying/shooting skills or the ultimate aircraft performance but team tactics and situational awareness. This aspect is currently highly lacking in Il-2, where it mostly ends in a furball lasting until one side is completely annihilated. In a scenario like this players feel like the AI is flying poorly, because our task is much easier than it historically was. It is very different compared to reality, if we "fly" a WW2 era aircraft using a comfortable joystick, not being subjected to any forces, combat fatigue, fear, heat, cold, noise, exhaustion and connection to the machine and its technical state. It is also quite different to survive a few QMB furballs and an ironman career with 150 combat missions. 

[SYN_Speck]

15 hours ago, Letka_13/Arrow_ said:

On the other hand, AI for offline flying in career should not represent some challenging combat bots that model online PvP behavior, which is more about an e-sport mentality.  I doubt that any sane WW2 pilot would try to get revenge after being hit, but rather bug out and try to make it back. 

Good AI in a historical combat sim is in my opinion more about using realistic tactics, which results in immersive experience. And historically it was more about survival, good positioning, using your tactical advantage to the maximum, short but intensive dogfights and generally realistic attrition rate with realistic chances of making it through by flying smartly. This is in my opinion not that much about exceptional flying/shooting skills or the ultimate aircraft performance but team tactics and situational awareness. This aspect is currently highly lacking in Il-2, where it mostly ends in a furball lasting until one side is completely annihilated. In a scenario like this players feel like the AI is flying poorly, because our task is much easier than it historically was. It is very different compared to reality, if we "fly" a WW2 era aircraft using a comfortable joystick, not being subjected to any forces, combat fatigue, fear, heat, cold, noise, exhaustion and connection to the machine and its technical state. It is also quite different to survive a few QMB furballs and an ironman career with 150 combat missions. 

 

 

That's a good summary of the issue. As an SP player now i just want to have an immersive flight sim experience, and the AI are the one element that just kill the immersion almost every time. It's not about their gunnery, although that is an element if it's ridiculously good or bad. It's about their consistently unrealistic behavior, especially in combat.

[[CPT]Crunch]

One huge problem would be the learning module should understand how us humans actually detect, visually track, identify, and than react, because when we fight we do it all based on visual response.  Pretty hard to get any learning going when the game as it is can't possibly track or understand our visual information fight.  The AI is all seeing or not and responds mainly to your direct controller inputs for the most part.  We're playing two completely different games with not much in common.

[IckyATLAS]

2 hours ago, [CPT]Crunch said:

One huge problem would be the learning module should understand how us humans actually detect, visually track, identify, and than react, because when we fight we do it all based on visual response.  Pretty hard to get any learning going when the game as it is can't possibly track or understand our visual information fight.  The AI is all seeing or not and responds mainly to your direct controller inputs for the most part.  We're playing two completely different games with not much in common.

I agree, but if you track in multiple scenarios and over many missions the behavior of a plane in various configurations during combat (air combat, ground attack), detailed trajectories (plane attitude  and speed, altitude, acceleration), firing, bombing etc. you can very well infer statistically a detailed pilot profile without going into the fuss of looking where eyeballs go, head rotation, hands actions etc. I have the feeling that the answer to direct inputs from the pilot senses is in the end not much important because all those fine details finally boil down to airplane maneuvering in space and firing. So analyzing that airplane behavior in a given context and repeatedly so as to have a statistical relevance will be enough and this can be learned by AI. You can have two pilots acting very differently in their cockpit in the sense of how they perceive their plane and environment but ending up with a similar profile, and this is correct as what counts is what the airplane does finally.

[Panzerlang]

I've flown the St Lo mission twice, in both cases two of the 190s crashed into each other just after taxiing and reaching the runway, in the first instance also wrecking me. I guess I'm not reporting anything new but when the AI can't even handle basic stuff like that it's easy to understand why its combat capabilities are so fubar.

I'd be interested to know how they're going to code the AI for the new game if they have nobody that can fix GB's.

[Aapje]

3 hours ago, [CPT]Crunch said:

One huge problem would be the learning module should understand how us humans actually detect, visually track, identify, and than react, because when we fight we do it all based on visual response.  Pretty hard to get any learning going when the game as it is can't possibly track or understand our visual information fight.  The AI is all seeing or not and responds mainly to your direct controller inputs for the most part.  We're playing two completely different games with not much in common.

A machine learning solution that gets taught to mimic human pilots can solve that, because it will just learn to ignore the information. So if the human player flying a certain plane cannot see an incoming fighter from 12 low, then the AI won't react even though you give it the information that the plane is there. After all, it is not trained to be the best, but to mimic human behavior.

 

Such an AI can also mimic situations where the pilot could see the plane, but fails to detect it, due to not checking his six, ground clutter, etc. For example, if only 40% of pilots would detect a plane coming from behind, even though they could spot it, then the AI would also only detect it 40% of the time. Such an AI would also mimic the pilot losing sight of the other plane.

 

Basically, a machine learning AI takes time to train, and to figure out what to train it with, but the upside is that all kinds of problems then solve themselves, like visibility issues and how well pilots look around them.

 

[Sobilak]

On 2/29/2024 at 8:22 PM, CzechTexan said:

I have also become more aware of unusual A/I behavior.  I do like the more aggressive head-ons.  However, the deadly sniper shots seem to be more frequent.  There are not even any rat-tat-tat hits on the aircraft to warn the pilot - just a single shot to the cockpit!  It's very frustrating.

I don't noticed that. Yes, after new update the AI accuracy is better in head on and at large angles, but about a some week of "fly" I never got hit in the head if I do maneuvered to don't be shoot. There was rat-tat-tat hits on the aircraft, but no hits to the head. The bomber gunners also missed head. Maybe I'm lucky.

[Panzerlang]

How hard would it be to make the AI target the engine/s instead of the pilot.

[AEthelraedUnraed]

On 3/2/2024 at 10:03 AM, giftgruen said:

This is an interesting discussion.

 

First of all: Yes. LLM is the way to go on long term to simulate intelligence in a way that we could call 'acceptable simulation of human intelligence'

However - a human pilots behaviour is not only controlled by knowledge and decision finding, a huan pilots behaviour is also ( largely ! ) influenced

by factors like mood, fear, braveness, feeling tired, beeing unconcentrated, beeing in anger, beeing cool ... and so on. Emotions. Character. 

 

So even if LLM could enable us to create bots with perfect intelligent behaviour from a 'technical abilities' point of view, this is not the

same like creating bots that simulate human pilot behaviour.

 

In my opinion, actual LLM ( though achieving a real breakthrough in AI ) is still far away from what we dream of.

Let alone having a LLM generated human behaviour simultanously for multiple bots, limited by a home computers CPU.

I think you're confusing terms. As 1PL-Husar-1Esk says, LLM stands for something completely different. The term you're looking for is probably CNN, Convolutional Neural Network. And as long as you have a way to model and train those factors you mention, there is no reason a CNN couldn't take those into account as well. Furthermore, running (which is different from training!) such a CNN can easily be done on even a low-spec home PC (usually on the GPU, not the CPU).

 

On 3/2/2024 at 1:32 PM, Aapje said:

For example, I think that it is likely that human pilots who get damaged are more likely to want to take revenge, or want to flee (fight or flight response). So if the damage level of the plane (and how long ago they were damaged) are used as parameters of the model, it can learn how human pilots react to being damaged.

 

The big benefit of a machine learning model over procedural AI is that the the former will automatically figure out most of these behaviors, once you feed it with the right data and enough data, while a procedural AI requires all of these behaviors to be hand-coded and hand-tuned.

The problem, as always with a neural net, is the training. To put it in laymen's terms, a neural net learns by example. To teach them how to react to being damaged, you'd have to *already* have examples of damaged planes and human reactions to that. These obviously have to be human pilots since if you train it using the current AI, all it does is learn to copy the current AI. While not impossible, there are a couple huge pitfalls with gathering human data:

- You need absolutely huge numbers of recorded dogfights. Something like 10,000 to begin with.

- You need to analyse this data and and turn it into chains of actions/reactions, other inputs (speed, altitude, proximity of other aircraft, ....) and results that your model could learn from. It is important that at this stage you already assign a value to all reactions depending on how good they are, meaning you'd need to come up with a method to objectively judge the worth of certain actions as well.

- If you train it against multiplayer humans, you're training it to copy multiplayer flying behaviour. Which is generally not what you want in SP; the reason I play SP is that it gives more historical gameplay.

 

The other option is what basically boils down to a sort of Generative Adversarial Network in which you pit two AI against each other and let them figure out what works well. This comes with similar problems of having to "rate" their behaviour as well as that you don't necessarily end up with realistic/historical human behaviour.

 

Another huge disadvantage, although this counts for AI in general, is that eventually you'd end up with an AI that does everything good. That's not what you want. Such an AI would be nearly unbeatable, or perhaps even avoid combat altogether when flying an inferior plane.

 

The difficulty of writing good game AI is not in having it make correct decisions, but in having it make believable incorrect ones.

 

4 hours ago, Aapje said:

A machine learning solution that gets taught to mimic human pilots can solve that, because it will just learn to ignore the information. So if the human player flying a certain plane cannot see an incoming fighter from 12 low, then the AI won't react even though you give it the information that the plane is there. After all, it is not trained to be the best, but to mimic human behavior.

 

Such an AI can also mimic situations where the pilot could see the plane, but fails to detect it, due to not checking his six, ground clutter, etc. For example, if only 40% of pilots would detect a plane coming from behind, even though they could spot it, then the AI would also only detect it 40% of the time. Such an AI would also mimic the pilot losing sight of the other plane.

 

Basically, a machine learning AI takes time to train, and to figure out what to train it with, but the upside is that all kinds of problems then solve themselves, like visibility issues and how well pilots look around them.

Things like this, you'd almost always want to model mathematically. An enemy is either visible or not depending on its relative position, which you can just calculate on a plane-by-plane basis. In the visible area, the odds of spotting it depend on where in the visible area it is. E.g. a real-life pilot might check its 12 o'clock every second, and its 5 o'clock every minute or so, which you can hard-code or use a RNG for. Having a Machine Learning AI learn these functions will tremendously slow down and complicate training, for little to no gain (it might even make things worse).

 

A misconception about IL2 is that it doesn't calculate visibility. It does. You can approach other aircraft from their dead angle just fine. The problem is that if you fly even slightly into their field of vision, they almost immediately spot you. IMO the pilots/gunners are much too attentive.

 

Same goes for night flying: you can usually approach bombers to within ~100m before they spot you. Unless you've accidentally left on your landing light, speaking from experience. I've also had AI lose track of me in the clouds, although it works a little quirky and seems to also depend on what mission logic is used.

[Sobilak]

4 hours ago, Hetzer-JG52 said:

How hard would it be to make the AI target the engine/s instead of the pilot.

But the pilot is more important target. If you shoot out the aircraft, but pilot bail out over his territory, then he take another aircraft and fight again with you or your camrades. It's absolutely logical move from AI.

Edited March 4, 2024 by Sobilak

[Gambit21]

1 hour ago, Sobilak said:

But the pilot is more important target. If you shoot out the aircraft, but pilot bail out over his territory, then he take another aircraft and fight again with you or your camrades. It's absolutely logical move from AI.

 

That's not really the point...at all.

 

 

[Aapje]

5 hours ago, AEthelraedUnraed said:

While not impossible, there are a couple huge pitfalls with gathering human data:

- You need absolutely huge numbers of recorded dogfights. Something like 10,000 to begin with.

- You need to analyse this data and and turn it into chains of actions/reactions, other inputs (speed, altitude, proximity of other aircraft, ....) and results that your model could learn from.

 

That's exactly why I suggesting starting sooner than later, because a lot of the time they'll wait for data to come in. And if they failed to record something important, they may have to start over collecting the data.

 

And yes, some data will need to be processed, especially from absolute values to relative values compared to the specific plane that the model is getting trained on at that moment. That said, I assume that they already need to do at least some of that for the procedural AI.

 

After you have a decent AI, it may also be possible to collect useful data from human pilots playing SP.

 

Quote

It is important that at this stage you already assign a value to all reactions depending on how good they are, meaning you'd need to come up with a method to objectively judge the worth of certain actions as well.

 

No, that is not needed if you train the model to mimic behavior. Then 'good' is behavior that matches human behavior.

 

5 hours ago, AEthelraedUnraed said:

- If you train it against multiplayer humans, you're training it to copy multiplayer flying behaviour. Which is generally not what you want in SP; the reason I play SP is that it gives more historical gameplay.

 

That is a good point, although as I argued, AI models can be directed to generate certain behaviors through input parameters. Although you do then have to be able to classify human flights/pilots in a way that correlates with the parameters.

 

For example, you can rate flights based on the behavior after damage. If they don't withdraw, but keep fighting, you rate them lower on the parameter value (and the higher the level of damage without a withdraw, the lower you rate them). Then IL-2 could offer the user the choice whether they want to fight planes that are more likely to withdraw. Presumably, some of these measured values would correlate with other behavior and you could use them to identify a certain kind of pilot (including more realistic pilots).

 

5 hours ago, AEthelraedUnraed said:

Things like this, you'd almost always want to model mathematically. An enemy is either visible or not depending on its relative position, which you can just calculate on a plane-by-plane basis. In the visible area, the odds of spotting it depend on where in the visible area it is. E.g. a real-life pilot might check its 12 o'clock every second, and its 5 o'clock every minute or so, which you can hard-code or use a RNG for. Having a Machine Learning AI learn these functions will tremendously slow down and complicate training, for little to no gain (it might even make things worse).

 

Except that real pilots obviously won't check the sky on such a regular interval like only a computer program could and potentially would. A real pilot is going to have target fixation, is going to suffer from being blinded by the sun, will lose planes in the clouds, will lose them in the ground clutter, etc.

 

Also, when losing a plane a human pilot won't just do something random, but will make an educated guess and will fly to a likely location and will focus on a region of the sky. So a very skilled pilot would then probably intentionally fly somewhere unexpected, while a less skilled pilot would fly more often to a likely location.

 

I don't believe for a second that is possible to model all this realistically using a procedural approach, especially if you also want to model realistic skill differences.

Edited March 4, 2024 by Aapje

[AEthelraedUnraed]

3 minutes ago, Aapje said:

That's exactly why I suggesting starting sooner than later, because a lot of the time they'll wait for data to come in. [...] And yes, some data will need to be processed, especially from absolute values to relative values compared to the specific plane that the model is getting trained on at that moment.

In this case, I don't think gathering the data is the issue, but annotating it. Converting it from absolute into relative is relatively easy actually, it's just mathematics that depend on the flight model. But that's not what I meant with my comment. What I meant is, let's say a Spit gets on the tail of a 109 with the same energy level. Now the 109 starts a climbing turn, and gets away. You should annotate this as a success, even if the 109 gets shot down by the same Spit 5 minutes later. The AI has to know whether a certain move was good or not, and *someone* has to judge this for basically every reaction that is taken. That's the whole idea of supervised Machine Learning

 

10 minutes ago, Aapje said:

No, that is not needed if you train the model to mimic behavior. Then 'good' is behavior that matches human behavior.

Even if your benchmark is "matches human behaviour" you still have to come up with some measure that assigns a certain value to this. Does it match "human behaviour" if 1/100 humans did this under similar circumstances? You need to come up with some numeric value to tell the AI if its actions were good or not. Regardless of whether this is according to the objective best or according to how well it matches human behaviour or not. You always need to quantify actions according to how "desired" they are in order to train the AI. And quite often, this quantification is the problem.

 

16 minutes ago, Aapje said:

That is a good point, although as I argued, AI models can be directed to generate certain behaviors through input parameters. Although you do then have to be able to classify human flights/pilots in a way that correlates with the parameters.

 

For example, you can rate flights based on the behavior after damage. If they don't withdraw, but keep fighting, you rate them lower on the parameter value (and the higher the level of damage without a withdraw, the lower you rate them). Then IL-2 could offer the user the choice whether they want to fight planes that are more likely to withdraw.

You're absolutely correct here, but it also massively complicates things if you're now suddenly differentiating between various different desired AI behaviours 

 

18 minutes ago, Aapje said:

Except that real pilots obviously won't check the sky on such a regular interval like only a computer program could and potentially would. A real pilot is going to have target fixation, is going to suffer from being blinded by the sun, will lose planes in the clouds, will lose them in the ground clutter, etc.

All of those are relatively easy to program using "traditional" programming. While for ML, you'd need to introduce several more input parameters which can massively complicate your model and slow down training.

 

22 minutes ago, Aapje said:

Also, when losing a plane a human pilot won't just do something random, but will make an educated guess and will fly to a likely location and will focus on a region of the sky.

True, and you *might* want to calculate that through a ML AI. But visibility itself is extremely easy to model mathematically/programmatically. Actually, even those educated guesses of where a plane may appear are pretty easy to describe mathematically; usually something like the position as well as its 1st and 2nd order derivatives are more than enough to obtain a good guess. No need for complicated, difficult to train and time-consuming Neural Nets here.

 

29 minutes ago, Aapje said:

I don't believe for a second that is possible to model all this realistically using a procedural approach, especially if you also want to model realistic skill differences.

You have a much more optimistic view of the human mind and the concept of "free will" than I have

 

Human behaviour is usually extremely predictable and easy to approximate using stochastic distributions. I fully agree that it is not possible to model human behaviour in its entirety using procedural programming, but there are very large portions of it that can easily be modeled that way without much "quality loss", if any. If there is any part of your network that you can do procedurally without much quality loss, this is almost always a good idea since this will lead to faster training and more predictable behaviour.

 

CNNs are incredibly powerful, but only if you know how to deal with them. This usually involves some specialisation of some kind. Even the human brain, the most versatile Neural Net there is, has various parts dedicated to some limited functionality. If you use one single Neural Net to do everything, you're often only complicating training while getting little in return. And while yes, you could train a Neural Net to calculate things like visibility, it's easier and faster (both training/programming time as well as CPU cycles) to just do that programmatically.

 

 

BTW, one thing I've wondered given your nickname; are you Dutch by any chance? ?

[Sobilak]

On 3/4/2024 at 10:24 AM, Gambit21 said:

That's not really the point...at all.

I  know. But in the game, when you crash land with heavy damage close to enemy AAA, the gunners continue to fire until they kill your pilot or you bail out of the aircraft. So AI is aimed at eliminating the pilot or forced him to bail out from aircraft.

[[CPT]Crunch]

Isn't the whole point of our flight sim hobby to recreate the WWII experience, that's not what this learning AI are going to get from the majority of us gamers.  Very few who actually have an inkling of what went on in the skies at that time, even fewer that will work as a combined force.  First thing out the window with that newly educated AI must be all the pretense.  Now you can have three dozen independently roaming free hunting Hartman's all looking to engage you from every angle all across the map who never miss their shot.  Good luck having fun with that berloga style free for all mess.  In the end that's what it usually boils down too when you allow us pilots into the design features.

[AEthelraedUnraed]

2 hours ago, Sobilak said:

I  know. But in the game, when you crash land with heavy damage close to enemy AAA, the gunners continue to fire until they kill your pilot or you bail out of the aircraft. So AI is aimed at eliminating the pilot or forced him to bail out from aircraft.

Erm no, the AI has no knowledge of the concept "pilot", let alone aiming for it.

 

The AI shoots at all enemy aircraft that are active. They are active as long as they have a living pilot. Which means that the AI shoots at all aircraft that have a pilot. But that's something else entirely than being "aimed at eliminating the pilot".

[Parazaine]

Have any of you played Wings over Flanders Fields or Wings over the Reich? The AI in those games is more realistic...they will kill you if you dont jink, but they dont magically snipe you as a poor substitute for decent AI.  They will act to avoid collisions. They can taxi and fly in formation competently AND they will flee combat if things are not in their favor, or start going badly.

 

Both games are based on an old engine,mscfs and whilst not as pretty as IL2 ,the AI and the SP campaigns are superb. Perhaps IL2 could learn something from them.

Also, yes,it seems like IL2 multiplayer is not in a good state currently. Some competent (as in well modeled) bots on the MP servers could go along way to flesh out empty MP servers.

Edited March 4, 2024 by Parazaine

[Majpalmer]

I loved WOFF. Yea, the AI was pretty good and the campaign game "Career" super. There are so many different types of aircraft. 

 

I switched when they went to the newer version a few years ago. For some reason, I couldn't get it to run om my new rig. 

 

The engine is MCFS3 from 2002. (Which I still have.). There are limits to how far they can go working with such an old core. Basically, the game is a glorified Mod. But it is great! 

 

Another issue I forgot about was that my copy of MCFS was a CD. My new rig didn't have a CD drive. 

[Sobilak]

7 hours ago, AEthelraedUnraed said:

Erm no, the AI has no knowledge of the concept "pilot", let alone aiming for it.

 

The AI shoots at all enemy aircraft that are active. They are active as long as they have a living pilot. Which means that the AI shoots at all aircraft that have a pilot. But that's something else entirely than being "aimed at eliminating the pilot".

And that's a point of problem of AI improved accuracy.

[Aapje]

13 hours ago, AEthelraedUnraed said:

The AI has to know whether a certain move was good or not, and *someone* has to judge this for basically every reaction that is taken. That's the whole idea of supervised Machine Learning

 

Again, this is not true if you teach the AI to mimic human behavior. Then the reward function is how closely it matches the real behavior. And secondly, you don't necessarily have to judge individual actions, since you would know all the actions that belong to the same pilot, so you could judge their entire flight record.

 

For example, you can judge their gunnery skills by measuring the hit percentage, or even the hit percentage based on relative position (so you can distinguish between pilots that are good at deflection shooting vs those than can only do easy shots). You can judge survival skills by how often they get shot down. You can just attacking skills in general by their kill percentages. You can judge (one form of) realistic attacking skill by the distance to the enemy from which they shoot on average. Etc, etc.

 

I would expect that if you find a few good measurements, you can also predict certain other behaviors that are hard to judge from the data, since people who conform to a 'type' in one way tend to conform to it in other ways.

 

13 hours ago, AEthelraedUnraed said:

Even if your benchmark is "matches human behaviour" you still have to come up with some measure that assigns a certain value to this. Does it match "human behaviour" if 1/100 humans did this under similar circumstances? You need to come up with some numeric value to tell the AI if its actions were good or not.

 

If you look at what LLMs do, then you see that they don't actually generate a single answer, but a probability distribution for various answers (well, tokens actually). They intentionally don't pick the most likely answer all the time, so the LLM becomes 'creative' instead of merely always producing the blandest, most predictable answer. This is actually an issue for factual questions, because the creativity then consists of wrong answers (known as hallucinating, although I consider that a bit of a misnomer).

 

An advantage of using a probability-generating AI for AI pilots is that there are no objectively correct behaviors if your goal is to mimic human pilots, although people in games especially can do some excessively dumb stuff. You can probably fix the worst issues with weird behaviors by simply only considering behaviors above a certain threshold of probability.

 

And then you can use a random number generator similar to how Dungeons & Dragons uses dice together with a probability chart. So the behavior that humans do 20% of the time then would also happen 20% of the time by the AI.

 

But of course you can play with this as well. For example, if the user prefers a predictable AI, you can weigh the more common actions more heavily or when the predictability slider is maxed out, only do the highest scoring action.

 

13 hours ago, AEthelraedUnraed said:

Regardless of whether this is according to the objective best or according to how well it matches human behaviour or not. You always need to quantify actions according to how "desired" they are in order to train the AI. And quite often, this quantification is the problem.

 

You're absolutely correct here, but it also massively complicates things if you're now suddenly differentiating between various different desired AI behaviours 

 

All of those are relatively easy to program using "traditional" programming. While for ML, you'd need to introduce several more input parameters which can massively complicate your model and slow down training.

 

I think that our fundamental disagreement is that your goal is to have a single AI that mimics a specific kind of pilot, while my goal is to have an AI that can be used to mimic different kinds of semi-realistic pilots. You are correct that it massive complicates things if you take the programmatic approach to do this, but a big strength of the AI approach (that uses the lessons learned from recent AI advances like LLMs), is that it can generate different behaviors organically based on the input parameters aka what you ask it to do.

 

I see people here demanding realistic AI pilots, but I strongly doubt that this is actually representative of all or most SP players. People often think they want something, but then actually reject it in reality. An example is that (AFAIK) all shooters, even the ones that aim for realism like Arma and Hell Let Loose, have sped up reloads, because people tend to greatly dislike long reloads while they are getting shot at.

 

I suspect that a very common reason why people prefer SP is because they don't want to fight against people with skill level 3000, even though it is realistic in reality that you might run into an ace. And I also expect that many people prefer an AI that starts spraying you relatively ineffectively from a larger distance, giving you a chance, rather than one that does what at least aces would typically do, which is to shoot from point blank range.

 

But with an AI that can produce different (semi-)realistic kinds of pilots, players could pick the kind of pilots they want to fight against. 1GCS could then experiment with the prompts for the AI to see what generates a type of pilot that the player may want to choose to fight against and then they could offer a very much simplified interface to the player.

 

And truly realistic AI seems to be impossible, since I don't think that we have any certainty about how pilots actually fought. They would have relatively basic training, as training wasn't very good back then compared to more modern times. Lots of training was on the job or by getting tips from experienced pilots.

 

13 hours ago, AEthelraedUnraed said:

BTW, one thing I've wondered given your nickname; are you Dutch by any chance? ?

 

Yes.