AI planes

[HazMatt]

I've set the enemy AI planes at different settings and it seems the only real difference between ace and novice are their gunnery accuracy. In my experience the plane all seem to try to zoom and rope you so if you have the better climbing plane it's very easy to beat the AI. Maybe I'm missing something? I've only been playing a couple months so it's quiet possible.

[Charon]

I believe AI level also affects the willingness of AI to break off from a ground target and defend itself, although this is both too low across the board, and the approach speeds for attackers are too low for it to matter much.

 

But yes, the AI is limited, especially in duels. In a campaign it becomes more about SA and making sure you don't get shot by an enemy you didn't see.

[HazMatt]

That makes sense. I was hoping maybe there was something to make the AI more of a challenge offline cause it feels like shooting fish in a barrel if you have a comparable plane. To make it a challenge I have to get a crappy plane and fight multiple AI and I'm not that good.

 

I have flow AH3 offline missions and the AI there seems to be much better. Maybe the devs could take a look at that AI and make some improvements? I find the ai there to be much more challenging.

 

One thing I've noticed with both the AI in IL2 and in AH3 is that they don't seem to bleed speed in turns or in the vertical as much as you would if you were in the same plane with say 20% fuel.

[BlitzPig_EL]

The AI are always in perfect trim.

[HazMatt]

Figures. lol

[Flying_Anchor]

13.02.2024 в 00:10, HazMatt сказал:

AH3

Sorry? What is it?

[LukeFF]

6 hours ago, Flying_Anchor said:

Sorry? What is it?

 

Probably Aces High 3.

[jojy47jojyrocks]

 

On 2/13/2024 at 1:40 AM, HazMatt said:

That makes sense. I was hoping maybe there was something to make the AI more of a challenge offline cause it feels like shooting fish in a barrel if you have a comparable plane. To make it a challenge I have to get a crappy plane and fight multiple AI and I'm not that good.

 

I have flow AH3 offline missions and the AI there seems to be much better. Maybe the devs could take a look at that AI and make some improvements? I find the ai there to be much more challenging.

 

One thing I've noticed with both the AI in IL2 and in AH3 is that they don't seem to bleed speed in turns or in the vertical as much as you would if you were in the same plane with say 20% fuel.

 

 

Starting to wonder if our own ordnance and fuel levels affect aircraft performance...

 

The AI in Offline just do a big zoom up and zoom down, and just keep doing that, rinse and repeat. It was ok before all this, they did not constantly go zoom up. I thought the AI improved, it seemed to get worse...

 

Edited February 23, 2024 by jojy47jojyrocks

[LukeFF]

6 minutes ago, jojy47jojyrocks said:

Starting to wonder if our own ordnance and fuel levels affect aircraft performance...

 

Yes, they always have.

[Gambit21]

On 2/12/2024 at 10:57 AM, HazMatt said:

I've set the enemy AI planes at different settings and it seems the only real difference between ace and novice are their gunnery accuracy. In my experience the plane all seem to try to zoom and rope you so if you have the better climbing plane it's very easy to beat the AI. Maybe I'm missing something? I've only been playing a couple months so it's quiet possible.

 

AI is in general unsophisticated with regard to differences between settings. You might surprised at exactly how unsophisticated in fact.

Hopefully this changes. 

AI in general will have to get much better if PTO is going to be a thing. 

[czech693]

You said the PTO word!

[taffy2jeffmorgan]

Every time there is an update I'm hoping to see that a vast improvement has been made with the AI in all aspects of the game, but no luck so far !!

[RNAS10_Mitchell]

Interesting,  my initial impression after the last update is that the AI in FC seem more aggressive,  and perhaps even more competent.   Further testing required though.

 

Update...  a bit more time flying against AI.   It's gunnery is too good now.  Long range sniper is unrealistic and not much fun.  This is the case even on veteran settings.   Needs to be dialed back.

Edited February 24, 2024 by RNAS10_Mitchell

[jojy47jojyrocks]

1 hour ago, RNAS10_Mitchell said:

Interesting,  my initial impression after the last update is that the AI in FC seem more aggressive,  and perhaps even more competent.   Further testing required though.

 

 

They just keep zooming up, most times, till stalling...

 

Ai used to be OK...somewhat.

 

Now, they just keep zooming up most of the time...

[357th_KW]

I've found that there are a lot of little things you can do in the mission editor that will significantly change how the AI react.  Likely what is occurring is that whatever game type you are playing, the opponents are all coded with the same set of instructions, and thus you get a similar result from a similar scenario.

[taffy2jeffmorgan]

another point which I have always advocated is being able to give and receive orders, I would have thought that by this stage in the games development we should be able to instruct our flight members to carry out tasks allotted to them, but then I suppose that part of AI my be an all new ball game and complicated to implement on a grand scale.

We shall just have to wait and see.

[LukeFF]

11 hours ago, jojy47jojyrocks said:

They just keep zooming up, most times, till stalling...

 

Ai used to be OK...somewhat.

 

Now, they just keep zooming up most of the time...

 

I can assure you they are doing much more than that.

Besides that, it's usually the lower-level AI skill levels that have fewer maneuvers they can perform. 

[EAF19_Marsh]

To be fair, they are closer to Big data, not AI: they never learn anything.

[R33GZ]

16 hours ago, taffy2jeffmorgan said:

another point which I have always advocated is being able to give and receive orders, I would have thought that by this stage in the games development we should be able to instruct our flight members to carry out tasks allotted to them, but then I suppose that part of AI my be an all new ball game and complicated to implement on a grand scale.

We shall just have to wait and see.

That is something I just can't get my head around.... because those were features we had in Il21946. Your flight was reasonably good at following commands. I just don't get it ?

[Yogiflight]

21 minutes ago, R33GZ said:

That is something I just can't get my head around.... because those were features we had in Il21946. Your flight was reasonably good at following commands. I just don't get it ?

They were better in that originally, but it got lost through the "AI improvements" over the years. Unfortunately the big AI revision, Jason promised years ago, when they hired a new AI programmer, never happened.

In my eyes one of the biggest chances for the new project is, they can write all that stuff from the beginning. The AI improvements in GB in the last years, always brought some worsenings with them.

[taffy2jeffmorgan]

6 hours ago, R33GZ said:

That is something I just can't get my head around.... because those were features we had in Il21946. Your flight was reasonably good at following commands. I just don't get it ?

Yes I remember that, you could even call your home field to get a compass bearing to find your way home and if you were flying for the Luftwaffe you would get that gruff German telling you, over and over again that you could not land landing,  " Nicht landing. Nicht landing

[LukeFF]

On 2/24/2024 at 4:01 PM, R33GZ said:

That is something I just can't get my head around.... because those were features we had in Il21946. Your flight was reasonably good at following commands. I just don't get it ?

 

Most of the developers from the original IL2 moved on to different projects/companies years ago.

[RNAS10_Mitchell]

@LukeFF, is there any difference in AI pilot abilities at the various settings (average,  veteran,  ace, etc) besides gunnery accuracy?   I've long suspected , that at least at the Veteran and Ace settings the only difference was gunnery accuracy.  Recent testing with average AI settings makes me curious if that observation is accurate. And if it so,  is also true for the other settings as well. Thanks in advance. 

Edited February 28, 2024 by RNAS10_Mitchell

[Stonehouse]

3 hours ago, RNAS10_Mitchell said:

AI pilot abilities at the various settings (average,  veteran,  ace, etc) besides gunnery accuracy

As far as fighter pilots go there is at least a few viewable differences. The max open fire range is one, aces in the stock game are at 800m and the other end novices are 400m. This alone explains why aces can seemingly 1 shot you from very long range. There are also differences in maneuver limits such as certain CAS values for different things that differ by skill as well as other things. Each skill has a different block of parameters for dogfighting/aerial combat. If you can use unGTP and unpack scripts.gtp and look at the folder data\luascripts\ai that results from unpacking the gtp file you can see these for yourself. Each aircraft has it's own file.

 

At lot of my/community interpretations are guesstimates but the ones above are fairly clear and can be proven by creating simple mods for a single aircraft and testing the results. I've never been able to find the configuration or control for fighter pilot accuracy by skill so I don't know whether it is simply the weapon dispersion and accuracy eg Hispano 20mm or M2 Browning etc or there are similar things to the gunner error parameters below where modders can reach. Gunnery accuracy (or what I believe is more correct, the degree of gunner error) as in AI in turrets on larger aircraft most definitely differ by skill. 

 

You can see the details for gunners in data\luascripts\worldobjects\bots (also results from unpacking the scripts.gtp file) and file turretcontrollerai.txt. My belief (which may be incorrect but matches what I have seen in long experimentation) is that a perfect shot is calculated for a gunner and then error via the turretcontrollerai.txt calc and dispersion is applied to this shot in 3 dimensions to get the resulting in game shot.

 

Edited February 28, 2024 by Stonehouse

[R33GZ]

3 hours ago, Stonehouse said:

As far as fighter pilots go there is at least a few viewable differences. The max open fire range is one, aces in the stock game are at 800m and the other end novices are 400m. This alone explains why aces can seemingly 1 shot you from very long range. There are also differences in maneuver limits such as certain CAS values for different things that differ by skill as well as other things. Each skill has a different block of parameters for dogfighting/aerial combat. If you can use unGTP and unpack scripts.gtp and look at the folder data\luascripts\ai that results from unpacking the gtp file you can see these for yourself. Each aircraft has it's own file.

 

At lot of my/community interpretations are guesstimates but the ones above are fairly clear and can be proven by creating simple mods for a single aircraft and testing the results. I've never been able to find the configuration or control for fighter pilot accuracy by skill so I don't know whether it is simply the weapon dispersion and accuracy eg Hispano 20mm or M2 Browning etc or there are similar things to the gunner error parameters below where modders can reach. Gunnery accuracy (or what I believe is more correct, the degree of gunner error) as in AI in turrets on larger aircraft most definitely differ by skill. 

 

You can see the details for gunners in data\luascripts\worldobjects\bots (also results from⁷ unpacking the scripts.gtp file) and file turretcontrollerai.txt. My belief (which may be incorrect but matches what I have seen in long experimentation) is that a perfect shot is calculated for a gunner and then error via the turretcontrollerai.txt calc and dispersion is applied to this shot in 3 dimensions to get the resulting in game shot.

 

That fairly well explains the sequence to firing and hitting, or missing, for AI (probably better described as bots), but what about the behavior of a friendly flight? In IL21946 you could, for example,  order your flight to 'attack enemy fighters' and then order your wingman to 'cover me' and that is exactly what would happen... creating 2 flights of 2. No way you could do that in GB unless it was scripted.

 

The old technology is surely owned by 1C, wether the previous developers are still in its employ or not, shouldn't matter. 

 

I can see that the current 'AI' behavior most probably came from RoF, but I feel it really is a blight on the IL2 series to know that the older game was/is better with respects to the behavior of friendly AI... or even AI in general.

[Stonehouse]

45 minutes ago, R33GZ said:

but what about the behavior of a friendly flight

None of that sort of stuff is viewable - at least as far as I know. I can only talk to what I have seen in regard to the actual gunnery side and some parameters and limits seemingly placed on fighter pilot AI in air to air. 

 

An example - when attacking bombers, aces will climb as they approach from astern until they have height advantage and then use it to accelerate faster into attack. One of the parameters in the aircraft text file in the data\luascripts\ai folder appears to control how much height is gained. Novice pilots will attack at the same altitude as the bombers and are therefore generally much slower in closing on the bombers and so suffer more. Each skill level can be given a different height advantage value. The larger the vector difference between bomber and fighter (speed of overtake is part of this) the larger the impact to the accuracy of the bomber gunners (bigger vector diff = bigger gunner error) and so it also impacts the success of the attack on the bombers as the fighters take less damage on the way in. 

 

However, the actual AI routines themselves and things like radio commands and how they are interpreted by AI is not available for our view along with a lot of other things understandably in a copyrighted product.  

 

Past that I don't have a comment beyond hoping that the dev team does have budget to rework AI over time. I hope it is sooner than later as things like radio commands to your flight are often problematic at best as you mention.

[[DBS]Tx_Tip]

Good calls and investigation Stonehouse. 

Understanding what, where, why and how the AI respond to MCU commands within different scenarios is a cornerstone of creating working and immersive content across all the venues BoX has.

[LukeFF]

9 hours ago, R33GZ said:

The old technology is surely owned by 1C, wether the previous developers are still in its employ or not, shouldn't matter. 

 

I can see that the current 'AI' behavior most probably came from RoF, but I feel it really is a blight on the IL2 series to know that the older game was/is better with respects to the behavior of friendly AI... or even AI in general.

 

The IP rights to the original IL2 were sold off to another publisher back in late 2021. Not that it really matters, because the code base is entirely different from what is being used now.

[IckyATLAS]

14 hours ago, LukeFF said:

The old technology is surely owned by 1C, wether the previous developers are still in its employ or not, shouldn't matter. 

Unfortunately it does matter a lot. In theory extremely well done. structured, written and documented and commented (conceptually, and down to detail) coding should be independent of its developer. In reality it is often not. Even using mission critical software and  like NASA, ESA, companies like Boeing, Airbus etc. it is never true, and only partly. This is why you have satellites crashing, planes falling etc.

I discussed with developers of various types in banking industry, Iphone app developers, scientific instrumentation developers, business office software developers etc. 

What you say here is Achilles heel in software development. Customers always find the cost of software development too high, don't want to pay, and so things are never done as they should be and nobody wants to pay just for documentation. The developer must do things in a hurry as the paid number of hours is minimal to the bare bone. The sales people often do not understand a word about code and coding so they sell projects at impossible prices for the developers. Result when the developers leave and new ones come in reusability of code is what it is. I have also discussed with developers in various branches that were given the task of reusing code left from previous developers that left and it was often a nightmare at various degrees. In some cases the code was considered useless. I mean to be used as is but could not be touched and to improve it it had to be rewritten. It is extreme but it happens.

The reason is that often there are multiple ways to write code solution to a problem and it is developer dependent. Often it is also for the developer a way to make himself a necessity not to be fired. So he has no incentive to write code so that he is not a key element of it. We are humans and so are things.

 

This is why the coming AI revolution in coding may well make code developers redundant and obsolete in the coming years. And as the boss of Nvidia said recently the aim of Nvidia and AI is indeed to free the people from having to code. It is an official target of their strategy. 

 

 

[Aapje]

@IckyATLAS

 

You are approaching understanding, yet you don't really get it.

 

Human ability to understand complexity is very limited and programming is often beyond our basic abilities, so we use simplifications, subconscious understandings and such to do things at the edge of our limitations. And what we can produce at that limit of our limitations is very much constrained by these very personalized ways of thinking that are not even just specific to the person, but even to the person at that specific time. This is why it is often hard for developers to understand their own code later on, let alone the code of others.

 

Documentation and comments cannot convey these intangibles and a big issue with them is that programmers often fail to update them when they change the code, so the documentation and the comments then become disinformation. This is a reason why writing more readable code is now preferred over adding lots of comments and documentation. 

 

Quote

 Customers always find the cost of software development too high, don't want to pay, and so things are never done as they should be and nobody wants to pay just for documentation.

It's more complicated than that. Developers are constantly learning and the demands keep changing, which means that the chosen solutions are never optimal. Ideally, you'd apply all these learnings and adaptations to new demands, to the existing code, but then you never actually get to deliver any features, because you spend all your time on updating the code. See Star Citizen for a project where a perfectionist programmer is in charge who keeps burning huge amounts of money rewriting things over and over to make it better. The game will never get finished, unless they change their management.

 

One of the main focus points of programmers over the last decades has been to make it easier to change existing code, at least for business cases (I don't know game coding). 

 

But it is always about finding a reasonable balance between delivering features and doing things right, insofar the latter actually exists. Because programming is not hard science, so the right way to do things is to a large extent subjective. And perfection doesn't exist, so it's always a choice for a specific set of advantages and disadvantages.

 

Quote

 Often it is also for the developer a way to make himself a necessity not to be fired.

For a very long time we've had developer shortages, so there is really no reason to do this. I think that nearly all of these accusations are from people who are dealing with code that was written in a way that makes sense to one developer, but not to another, and then they assume that it was written in an intentionally hard to understand way, when the reality is that this made the most sense to the developer who wrote it.

 

Quote

This is why the coming AI revolution in coding may well make code developers redundant and obsolete in the coming years. And as the boss of Nvidia said recently the aim of Nvidia and AI is indeed to free the people from having to code. It is an official target of their strategy. 

The history of programming is full of innovations that were argued to end the need for programmers, but somehow the demand for programmers just keeps going up.

 

For example, reusing code was supposed to make most programmers obsolete, but instead of doing the same thing with fewer people, we just started to make much more complex and advanced programs. So instead of the complexity of writing all the code ourselves, it became the complexity of figuring out how we can achieve things by using existing code in the right way. The job changed, but it never went away. On the contrary, because programmers could now deliver more value with the same work hours, software became viable for more things, so the demand for programmers actually went up.

 

The only way I see AI making programmers obsolete (rather than just changing their jobs), is if AI provides a level of general intelligence that equals or beats humans across the board. Even in that case, I think that programmers have a much better shot to either stay employed (in a technical capacity) than most other employees, or to be able to stay alive in the underground resistance.

Edited February 29, 2024 by Aapje

[EAF19_Marsh]

6 minutes ago, Aapje said:

The only way I see AI making programmers obsolete (rather than just changing their jobs), is if AI provides a level of general intelligence that equals or beats humans across the board.

 

There is no common definition of ‘intelligence’: it can mean a lot of things and is usually used to mean ‘not obviously stupid’. Which is, of itself, a concept that is also difficult to pin down. Most ‘AI’ is not ‘intelligent’, it simply has a huge repository of data. Since when did any computer-controlled aircraft in this sim actually show ‘intelligence’? It does not and it can not.

 

But I am sure all these grand predictions of ‘AI’ and computers being the amelioration of mankind are accurate. Said no one that bas ever used Excel.

 

But, hey, I have a min-nuclear powered house as forecast in the ‘50s and work in a paperless office as promised in the ‘90s. Oh, hang on…

[Aapje]

On 2/28/2024 at 4:14 AM, Stonehouse said:

As far as fighter pilots go there is at least a few viewable differences. The max open fire range is one, aces in the stock game are at 800m and the other end novices are 400m. This alone explains why aces can seemingly 1 shot you from very long range.

This seems very different from historic reality. As far as I know, real ace behavior was to shoot from fairly short ranges, while novices would spray and pray from long distances.

 

And the biggest differences to actual combat success and survivability were probably not primarily the shooting, but in creating advantageous attack situations for yourself and either disengaging when the odds are not so good, or forcing the enemy into endangering himself. For example, when you are under attack, forcing the opponent to deplete their energy state to stay in their advantageous attack position, which means that they in turn become very vulnerable to attackers with a better energy state.

 

This is probably why aces would then have much better survivability. If they were under attack by a rookie, they could probably outmaneuver them or escape. Or the target fixated rookie would be attacked by a friendly. If they were under attack by an enemy ace, they could create a situation where the attacking ace would decide that it was too risky to deplete their energy state to kill the ace they were attacking, even though they could.

 

So I would think that the different levels of combat AI should ideally reflect all of this and more. In fact, I would like there to be different skill levels for various skills and traits.

 

For example, a pilot can be an expert at keeping his energy level high, or have a tendency to pull to hard or to pull the throttle rather than climbing to keep his energy state high. A pilot can be great at flying close to the stall limits of the plane or not so good at it. And then depending on their trait, more novice pilots either don't use the limits of their plane, or they are aggressive and are more likely to stall out or such. A pilot can constantly looking around, or they can never check their six and not scan around them that often. A pilot can have good eye sight and spot far away or have bad eye sight and spot planes only when they get closer. And of course spotting against ground clutter should be harder for any AI pilot. Target fixation levels should be different between pilots, where some don't even disengage when they themselves are attacked, while others disengage when things get too risky. I could come up with many more of these, including deflection shooting skill.

 

Ideally, AI pilots should then vary in skill and traits on all of these, where the higher AI levels simply have higher stats on average. But at any level you would find pilots with different strengths and weaknesses.

 

Of course, I don't know to what extent this is actually possible, but making 'aces' snipers that PK you from a large distance is the opposite of realism and for most people probably the opposite of fun.

[Aapje]

1 hour ago, EAF19_Marsh said:

There is no common definition of ‘intelligence’: it can mean a lot of things and is usually used to mean ‘not obviously stupid’

Yes and no. When we compare people with each other, we can measure their IQ, which fits the criteria of being a rather good way to measure things:

- It is quite stable across tests and time

- It is very predictive on the group level, so we can make fairly good predictions about the average ability of people with a certain IQ to finish certain levels of education, do certain jobs, do certain tasks, etc.

 

The issue with computer intelligence is that computers don't operate in the same way that we do, so where we need a general level of intelligence to do certain tasks, the computer can instead be a 'master of one'. For example, some computer scientists once believed that only a highly intelligent computer could do chess, but it turns out that a computer can beat the very best humans with a computer program that is very good at chess and not good at anything else.

 

So when you are comparing human intelligence to computers, it is always going to be an apples to oranges comparison, unless you manage to recreate the human brain in a computer (which some people try, but we are very far from making that possible). You cannot look at what the computer is doing internally to know whether it is intelligent. Of course, the same is true for people. We cannot distinguish between a smart and dumb person by dissecting their brain or measuring their brain, but merely by looking at what they can actually do.

 

This is why Turing came up with the Turing test, where he argued that it doesn't really matter what the computer really does. If they can talk/reason/etc in a way that is indistinguishable from real people, then they have at least the same intelligence abilities as people. The current LLM (large language model) AI systems like ChatGPT definitely are way closer to this ideal than anything that came before, although they are clearly not there yet.

 

Is that approach going to reach the level of general intelligence? People much smarter than me are divided on the topic. At the very least, it seems to be a question that is too difficult to know the answer to. Is it going to match or beat (many) people at tasks that we now consider dependent on human intelligence? Yes, it already does.

 

Quote

Most ‘AI’ is not ‘intelligent’, it simply has a huge repository of data.

Not really. LLMs are a combination of data with relationships between that data. For example, they encode that 'computer' and 'science' are words that are more closely related than 'computer' and 'poop'. This allows it to write sensible prose that are not just random words, but sentences that make sense.

 

This is actually also a key feature of the human mind. Our brain is a large network of neurons that have stronger and weaker links to each other. When you learn things, this network changes to encode knowledge within it, just like the LLM model changes the relationship between words when you train that network.

 

It is remarkable that the LLMs can do the same trick that human brains do, where we can learn all kinds of things that are somehow encoded in a network of a huge number of relatively simple neurons, where it is really the specific relationship between billions of neurons that allow us to learn, memorize, speak, etc. Somehow this allows me to write these sentences, even though none of the sentences I write seem to be stored in a specific location in the brain. They are somehow encoded in the relationship between the neurons.

 

That LLMs can do this same thing, where the sentences they write are also not specifically stored, but somehow encoded in the network, is very exciting, because it is the closest we've come to making computers do what human brains really do, rather than make something that produces somewhat similar results, but works completely differently.

 

Of course, the LLMs still use different technology than the human brain, with the details being very different. The question is to what extent those differences limit the LLMs or make them more capable. And of course to what extent it makes them different from the human brain.

 

I personally suspect that the human brain is much more fine-grained than the current LLMs. On the other hand, we can make the LLMs do way more learning. Also, our brain is amazingly efficient where it uses only 20 watts, but we are also limited by that. The LLMs are trained on systems using enormous amounts of power. So are these differences going to result in an AI that is limited to capabilities less than (most) people, will it equal us, will it beat us? Whose to say? But I feel that this is the first time in history where a computer beating people in general intelligence is at least a plausible outcome if we develop the technology we have further.

Edited February 29, 2024 by Aapje

[EAF19_Marsh]

18 minutes ago, Aapje said:

Yes and no. When we compare people with each other, we can measure their IQ, which fits the criteria of being a rather good way to measure things:

- It is quite stable across tests and time

- It is very predictive on the group level, so we can make fairly good predictions about the average ability of people with a certain IQ to finish certain levels of education, do certain jobs, do certain tasks, etc.

This is an extremely long and interesting debate that I would enjoy over a beer or 5, but I politely disagree with you and do not have the time to type out. Not personal at you. ?

 

Yes, IQ is a shorthand and is good enough as a rough guide. Yes, the Turing test is a measure of ability to mimic a human. Yes, undoubtedly increasingly sophisticated progress will become (worryingly) close to homo sapiens’ mental capabilities. That brings the good and the bad.

 

But run a QMB, shoot down the computer aircraft, run it again and it will make exactly the same errors. Pull the same maneuver in the same dogfight and it will make exactly the same errors. It has a data repertoire, but no ‘intelligence’.

 

The team did a passable job, but within current limits the code does not have the capacity for ‘intelligence’.

[RNAS10_Mitchell]

19 hours ago, LukeFF said:

 

The IP rights to the original IL2 were sold off to another publisher back in late 2021. Not that it really matters, because the code base is entirely different from what is being used now.

Anyone know if "another publisher" is doing or planning on doing anything with it?   Or who "another publisher " might actually be?

Edited February 29, 2024 by RNAS10_Mitchell

[Aapje]

1 minute ago, EAF19_Marsh said:

But run a QMB, shoot down the computer aircraft, run it again and it will make exactly the same errors. Pull the same maneuver in the same dogfight and it will make exactly the same errors. It has a data repertoire, but no ‘intelligence’.

Yes, but the current code is not using the LLM approach, but the old approach of 'let's encode some explicit rules and behavior that we think mimics human behavior.' That is exactly the approach that didn't make much progress for decades when it comes to creating 'intelligent' computers and is now soundly beaten by the LLM solutions. I think that the future of flight sim AI is going to be the LLM solution, where the computer learns from the behavior of human pilots.

 

If you look at my very incomplete explanation of what a proper AI pilot should do in my opinion, then it is probably way too complex to code this all out explicitly, so only an LLM solution can really do a good job.

 

I think that this should work:

1. Collect a ton of data from multiplayer, like all the control inputs of all pilots, the locations & speed of the planes, the type of plane, ground targets locations, etc. Also keep track of the k/d results (or even better a skill rating similar to how chess does it) to classify pilots by skill level and by risk taking behavior.
2. Distill this data into learning data. This would primarily consist of simplifying the data to only the things that matter. For example, two planes on opposite sides of the map have no impact on each other, so when you try to teach the system to learn from the behavior of the human pilot in plane 1, you don't have to give it information about the plane on the other side of the map. But if two planes are close together, then the pilot of the plane will often react to the other plane, so an AI also needs to be able to learn how to react, for which it needs to know what plane 2 did.
3. Teach a model using the gap method. This means that you leave out some of the control inputs from the plane whose behavior you are teaching the model to mimic. Then you make the model guess what control input the human pilot gave. You then reward the AI when it makes the right guess and punish it for making the wrong guess. Based on the reward/punishment it will then update the model to more often give that input at that time.
   Using this method is effectively the same as giving humans a math problem where they have to figure out what the value of X is in an equation, so it is a classic learning method. However, where it takes a person a long time to solve such a math problem, the computer can solve a ton of such problems in a fairly short time and thus learns very rapidly.
4. Put the model into the game and run it locally for each AI pilot. You can then tell the AI that this pilot has a certain skill rating and risk taking behavior (and perhaps some other behaviors that you put into the model). The AI will then generate the control inputs for the plane that mimic what a human pilot would do of that skill level, with that risk taking behavior, etc.

 

The above can be made better by including things like communication, so then the AI pilot can react to messages from human or other AI pilots, as well as send it's own messages. So it can tell you that there is a plane on your six. And also some form of memory, which you can for example use to have the AI pilot follow orders. So an AI pilot with orders to attack a bomber formation will ignore ground targets, but when the goal is to do CAS, it will attack ground targets.

 

The difficulty with having 1GCS or another flight sim company create this is presumably:

• You need someone with the knowledge & skills to build a custom learning model. The trick is to make it as complex as needed, but not too complex, so it becomes too expensive to teach the model and run the model. Of course, right now there is a huge lack of people with LLM skills & knowledge.
• It costs money to do the learning phase. Cloud providers are working on providing systems that you can use for a fee, but it is early days for this
• Gamers need to be able to run the model locally. The CPU companies are working on putting this into their processors, but it will take a while for people to have this.

But eventually this seems like the logical path for sims to take, with the emphasis on eventually.

[LukeFF]

4 hours ago, RNAS10_Mitchell said:

Or who "another publisher " might actually be?

 

Fulqrum Publishing

[1PL-Husar-1Esk]

4 hours ago, Aapje said:

I think that the future of flight sim AI is going to be the LLM solution, where the computer learns from the behavior of human pilots.

To do that you need a complete product, so first is a game without AI. Then AI is learning from players,  then you  put AI inside the game.

Since 80 % sales are coming from Single player gameplay this very risky business.

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

[RNAS10_Mitchell]

6 hours ago, LukeFF said:

Fulqrum Publishing

 

Interesting.   Thank you!  I see they have an add-on for cliffs of dover.  

[Aapje]

16 minutes ago, 1PL-Husar-1Esk said:

To do that you need a complete product, so first is a game without AI. Then AI is learning from players, then you put AI inside the game.

Since 80 % sales are coming from Single player gameplay this very risky business.

Both Il-2 and DCS are already products designed primarily for multi-player, which don't provide an optimal single player experience. For IL-2, we even have a third party tool to generate campaigns and there is a lot of complaining about the AI. And lots of single player features came quite late to BOX.

 

I personally think that game companies like 1GCS are already playing a dangerous game by focusing so much on MP.

 

So I think that having the multiplayer gamers generate the data for the AI model would be great for combat flight games, since it would create a much improved experience for single player gamers, but would also be a big boon for the multiplayer gamers. We've seen a lot of people complaining about empty MP servers lately, but with a strong AI, you can always give multiplayer gamers a good experience too. When the servers are empty of human pilots, you just add more AI planes, that don't conspicuously stand out by having unrealistic behavior, since they would be trained to act like real people.

 

Multiplayer games without good bots tend to have this threshold where once the player base drops too low and people can't find online servers with enough players, more players quit, so the servers get even more empty, etc. With good AI players, game companies can ensure that the servers are always full.

 

So the game would improve for both single and multiplayer gamers & the game company has a more stable income, since the MP gamers don't suddenly run away all at the same time when the game becomes a bit less popular and the servers have empty periods.