The AI thread

[kodamcity]

4 hours ago, BraveSirRobin said:

It takes years to train a fighter pilot. Then it takes years of experience before that pilot is actually really good.  We’re playing the game on a desktop computer. The reason the AI isn’t very good is pretty obvious.

Yes, training pilots takes a very long time. Even 17 years isn't enough to execute a simple loop or Split S maneuver...

[BraveSirRobin]

24 minutes ago, kodamcity said:

Yes, training pilots takes a very long time. Even 17 years isn't enough to execute a simple loop or Split S maneuver...

If you attempt a loop in a fight you’re probably going to die.  Maybe the AI is better than you think it is.

[kodamcity]

5 minutes ago, BraveSirRobin said:

If you attempt a loop in a fight you’re probably going to die.  Maybe the AI is better than you think it is.

That's nonsense regarding my comment. You know exactly what my intention was, yet you are blurring the point. My intention was to address the fact that they haven't been able to perform even simple defensive maneuvers for 17 years.

Still, it is an accurate point. However, that does not mean AI aircraft should never choose loops as a defensive maneuver. It depends on the situation, and it is a maneuver that can certainly be attempted if one lacks experience.

Conversely, is it a truly correct defensive maneuver to turn in only one direction and occasionally execute -G maneuvers, as is currently done? Do you agree with my opinion regarding an AI that has been unable to execute a Split S as a defensive maneuver for 17 years?

I do not want an AI that performs defensive maneuvers based on perfect theory. (In fact, the defensive maneuvers the AI is currently performing are by no means based on perfect theory.) I want to see an AI that alternates between left and right turns, performs loops, dives, and performs insane roll maneuvers to survive. We need to see them make mistakes from time to time to feel like we are fighting against a human. Turning unconditionally in only one direction like this is clearly wrong.

[BraveSirRobin]

So AI that is good, but not too good.  Difficult to fight against , but not so difficult that you can’t win (seriously, no one wants AI that always wins). Random bad decisions, but not so bad that they’re too easy to kill.  Sounds easy.  Total mystery why no flight sim has ever managed to do it.

 

Oh, and the AI has to be able to act like this no matter how skilled (or unskilled) the human opponent is.  Piece of cake.

Edited June 20 by BraveSirRobin

[kodamcity]

15 minutes ago, BraveSirRobin said:

So AI that is good, but not too good.  Difficult to fight against , but not so difficult that you can’t win (seriously, no one wants AI that always wins). Random bad decisions, but not so bad that they’re too easy to kill.  Sounds easy.  Total mystery why no flight sim has ever managed to do it.

 

Oh, and the AI has to be able to act like this no matter how skilled (or unskilled) the human opponent is.  Piece of cake.

Edited 8 minutes ago by BraveSirRobin

Sigh... Have you seen the WOFF gameplay video? Have you tried playing IL-2 1946? At this level, other users, including myself, would not raise any complaints about the AI level.

Furthermore, I wouldn't complain even if the AI level were just on par with the Strike Fighters series.

 

 

Edited June 20 by kodamcity

[BraveSirRobin]

Just now, kodamcity said:

Sigh... Have you seen the WOFF gameplay video? Have you tried playing IL-2 1946? At this level, other users, including myself, would not raise any complaints about the AI level.

Yes, someone posted a video of an SE5 turn fighting with a Dr1 on the deck as evidence of WOFF’s excellent AI.  I have ignored all claims of excellent WOFF AI ever since then.  
 

IL-2 1946 AI cheated.  It used a different flight model from the human player.  It was actually far more infuriating than anything that I have ever seen in GB.

 

sigh…

[kodamcity]

2 minutes ago, BraveSirRobin said:

Yes, someone posted a video of an SE5 turn fighting with a Dr1 on the deck as evidence of WOFF’s excellent AI.  I have ignored all claims of excellent WOFF AI ever since then.  
 

IL-2 1946 AI cheated.  It used a different flight model from the human player.  It was actually far more infuriating than anything that I have ever seen in GB.

 

sigh…

Good heavens. You keep misinterpreting what I'm saying.

Then, is it right for the AI in RoF, GB, and Flying Circus to only turn in one direction as a defensive maneuver?

Even if the AI in WOFF, IL-2 1946, and Strike Fighters cheats, I think it is much more immersive from the opponent's perspective.

 

 

Even if the AIs in those games cheat, players are fully capable of dealing with them. They may be difficult to deal with, but they are not invincible.

[kodamcity]

9 minutes ago, BraveSirRobin said:

Maybe for you.  I thought it sucked trying to fight UFO’s. Complete and total immersion killer.

Yes, that is correct. You can complain about those cheating AIs. After all, they aren't perfect either.

In that case, for me, dealing with AIs that only turn to one side in defensive maneuvers for 17 years was the worst. You seem to think this AI is far superior to them.

And you consider my argument that the AI needs to be improved to be an impossible demand.

Edited June 20 by kodamcity

[BraveSirRobin]

13 minutes ago, kodamcity said:

You seem to think this AI is far superior to them.

Nope.  Just a lot less infuriating. In GB an easy kill is consistently an easy kill.  In IL2 an easy kill frequently turned into an ammo draining slog.  Infuriating. 

16 minutes ago, kodamcity said:

And you consider my argument that the AI needs to be improved to be an impossible demand.

I think the problem is that improvement is ignored because the AI still won't be good enough and never will be.

[kodamcity]

Try again. Unless the AI is set to Ace Pilot difficulty, you can shoot it down quite easily in IL-2 1946 as well. The Ace pilots demonstrated true ace capabilities. And it was fun to take on the challenge. However, in GB, while the flight skills remained the same, only the shooting accuracy was boosted to an absurd level. This is where people like me are raising complaints.

I admit, just like you, that even if the AI improves, complaints will continue endlessly. It is impossible to satisfy everyone. However, as I said, the defensive behavior of the AI currently seen in the GB series and the Korea Early Access must be improved.

Even if we simply add a feature that alternates between turning left and right from the current stage where the AI only turns in one direction, it will feel more human than it does now. It will still be easy to shoot down the AI. Adding this variable won't be that difficult.

[BraveSirRobin]

4 minutes ago, kodamcity said:

Try again. Unless the AI is set to Ace Pilot difficulty, you can shoot it down quite easily in IL-2 1946 as well. The Ace pilots demonstrated true ace capabilities. 

I always set it to Ace and it did nothing of the sort.  Nothing but UFO bs. 

 

 

Edited June 20 by BraveSirRobin

[kodamcity]

1 minute ago, BraveSirRobin said:

I always set it to Ace and it did nothing of the sort.  Nothing but UFO bs. 

As I mentioned, then set the AI skill to be lower than Ace difficulty. You can shoot it down just as easily as in the GB series. And you are distorting my argument again.

[BraveSirRobin]

6 minutes ago, kodamcity said:

As I mentioned, then set the AI skill to be lower than Ace difficulty. You can shoot it down just as easily as in the GB series. And you are distorting my argument again.

I played enough 46 to know that the AI was crap.  I don't need to relive it.

[Leifr]

9 hours ago, LuftManu said:

Hey guys! Glad you enjoyed my video and others.

I have been doing some AI experiments ( @Leifrcan tell) and there is some great improvements. Need more time to show too 🙂

Also, from Command T. 

 

 

This is a first look at the fligh model and force feedback implementation of the new IL2 KOREA combat flight simulator. Overall, I'm impressed!

This is an excellent video.

I'm starting to be of the opinion that a lot of the content creators who have received an early access copy, and released videos of Korea, have been playing with the player-assists and rookie AI levels set to default. The flight characteristics demonstrated in CommandT's video look much more refined than others I've seen, and it appears to be because he isn't using the game assists. Also, thanks for the B29 AI demonstration @LuftManu; again, folk are using 'Novice' AI as a setting and it's obviously grossly impacting the gunners, having observed otherwise with a higher difficulty skill, I am more convinced.

I suppose a lot of content creators have simply cracked open the game and started playing without too much thought as to what they're showing, and as such I will appreciate more videos produced by folk such as yourself Luft and the more conclusive reports made by CommandT above.

 

S!

[KC23]

4 hours ago, BraveSirRobin said:

So AI that is good, but not too good.  Difficult to fight against , but not so difficult that you can’t win (seriously, no one wants AI that always wins). Random bad decisions, but not so bad that they’re too easy to kill.  Sounds easy.  Total mystery why no flight sim has ever managed to do it.

 

Oh, and the AI has to be able to act like this no matter how skilled (or unskilled) the human opponent is.  Piece of cake.

I thought Rowan's Mig Alley hit pretty close to that AI mark. It also had one of my all-time favorite true dynamic campaigns.

[AEthelraedUnraed]

6 hours ago, BraveSirRobin said:

I think the problem is that improvement is ignored because the AI still won't be good enough and never will be.

I beg to differ.

The issue is not really making AI that is good, as much as AI that is believably bad. It's perfectly possible, and in fact not even extraordinary hard, to develop AI that always makes the best decisions. However, that would result in AI that's not always as fun to play against. For instance, an inferior enemy aircraft would always flee as soon as you're spotted.

In fact, some of the issues with the current AI stem from the fact that it's already too good. E.g. the spotting system and situational awareness.

[BraveSirRobin]

2 hours ago, AEthelraedUnraed said:

I beg to differ.

The issue is not really making AI that is good, as much as AI that is believably bad. It's perfectly possible, and in fact not even extraordinary hard, to develop AI that always makes the best decisions. However, that would result in AI that's not always as fun to play against. For instance, an inferior enemy aircraft would always flee as soon as you're spotted.

In fact, some of the issues with the current AI stem from the fact that it's already too good. E.g. the spotting system and situational awareness.

When I say “good enough” I mean “acts human”.  The only area that I would expect any improvement is AI that has relatively human spotting/awareness.  But even that is probably extremely difficult.  Everything else will be obvious bot, probably forever.

[MajorMagee]

Here's an example of why getting the AI solution "right" is very complex.

Quote

AI OVERHAUL

Seeing an update come out without a single vehicle or mission may seem unusual, but we have a very good reason for this. (Aside from the fact that the majority of the team's efforts are pouring into the remaining milestones on the game's roadmap, which are complex and deserve ample time to build and refine!)

"Content" can only get you so far; the underlying game systems are what carry the experience. For that reason, we decided this time around to put a strike team on one of the issues that's been on our plates for a while: improving the AI crew performance. This is both a longstanding internal goal for us and a popular player request (enough so that it made it into the Frequently Asked Questions page).

The idea of "improving the AI" is too multifaceted to properly get into here, and even the extensive changes we've made in this update are just scratching the surface of what we're working on, but it's a solid start in two main areas: simulated vision and simulated awareness. Both are worth explaining in more detail.

COMMUNICATION AND AWARENESS SYSTEMS

Enemies now use their radios

Perhaps the most basic issue with GHPC's previous AI system was often summed up as: "They lack object permanence." Even backing down a hill for a few seconds was enough to make the enemy gunners give up trying to hit you and resume scanning the horizon like nothing had happened. This very obviously needed to be remedied at some point.

Now, the AI crews have full access to the radio system - the one that's used to send enemy location pings to your tactical map. The exact same communication network simulation is running for the enemy team, too, and now they have the ability to respond to what they know! Platoons will assign a portion of their members to watch for nearby reports, anticipating the detected enemy making an appearance soon. They'll keep watch until enough time has passed that they feel their attention is needed in their original assigned sectors, at which point they'll return to standard mode. But if they hear about a new, closer threat, they'll orient on that one instead.

What this means is that if you are flanking a position and get spotted by the scouts (and if they survive long enough to send a report on the net), you can expect every enemy formation nearby to start watching for you once the information filters back to them. If you approach from the same angle you were reported at, you'll find that they have guns up and will not take long to spot and engage you. To keep the upper hand, you'll need to either reposition frequently or never get spotted long enough for the alert to go out!

Enemies are more coordinated within a platoon

If you engage a group of enemy vehicles and then decide to back off to reload, they'll no longer lose interest in you immediately. Instead, some of the members will watch your last position, waiting for you to re-appear. It should be quite a bit harder to play "peekaboo" with a defensive formation and expect easy kills. In fact, appearing in the same place will not only get you seen quickly but also cut down on their re-aim time! (More on that below.)

There's also been a system in the game for some time that allows platoons to react to one of their members getting hit. If a vehicle survives a shot, the crew guess at the location of the shooter, turn the gun toward it, and (if still alive long enough) alert the rest of the platoon to follow suit when they can. We've fixed some issues with this system, and it should nicely augment the other communication enhancements.

Friendlies report what they see

For some time, we've had a way to see reported enemy locations on the map, with a walkie talkie icon that popped up to let you know a new message had arrived. We found that this system was both vague and unintuitive, with many players asking what the icon means and never actually checking the map. We've now replaced the icon with a full text message, explaining who saw what. The map pings are still present, but you can get a sense of their urgency by simply reading the text message.

As a bonus, we've also patched Close Air Support (CAS) pilots in to this radio message system. Now, any active CAS flight will report its status to you, including what target it's decided to strike. We've seen some confusion about whether CAS planes do anything at all, despite their bomb runs being one of the most powerful "win buttons" in the game - so we hope this system will help clarify whether CAS found a target to bomb and provide a hint when it wasn't the one you expected.

VISION SYSTEMS

Once a crew is aware of a target, they still need to acquire it. Previously in GHPC, this process was heavily abstracted, amounting to a lightly randomized countdown timer that (in theory) responded to the target's distance and speed, the night vision systems, light conditions, and the tuning setup of the shooter's tank type. In practice, for various reasons, this did not work the way we intended. Not only was the system too basic to properly convey the subtleties of different tanks' performance beyond night vision effects; the actual timing of target spotting was generally homogenized and flat. We decided to do something about all of this.

Vehicle visibility score

With this update, units in GHPC are now rated on how visible they currently are, which heavily affects how much difficulty an observer will have in finding them. The list of things that affect vehicle visibility are currently:

(New)

• How large or obvious the vehicle is inherently

• Whether the vehicle is in partial defilade ("hull-down") from the observer's perspective

• Recent weapon fire by the vehicle

(Previously broken, now working)

• The vehicle's speed

• The vehicle's distance from the observer

(Existing)

• Whether the vehicle is casting an infrared light beam (only in cases where the observer is using near-IR night vision at night)

• Whether the vehicle is in an illuminated zone from a flare (only in cases where the observer is using non-thermal optics at night)

If a vehicle's visibility is being boosted by firing a weapon, the strength of the effect depends on what kind of weapon it is. The effect wears off with time, but you can also dodge it faster by leaving the area where you fired the weapon.

Spotter vision score

When a crew is scanning for targets, their success is now influenced not only by the visibility of the targets but also by the vision abilities of their own equipment. Previously, we only simulated the effects of night vision optic types on sight ranges at night or in rain. Now we also adjust spotting times based on:

• The magnification quality of the gunsights

• Whether there is a thermal sight available for observation

Much like the rest of the new vision features, this score is now standardized across all vehicles and emplacements in the game (i.e., if a Soviet vehicle is rated as having a high-zoom, non-thermal sight, it will have the same spotting speed as a West German vehicle with a high-zoom, non-thermal sight). This is much easier for us to set up and tune, and it also helps avoid accidentally favoring one vehicle unfairly.

Nearby target detection

A rather amusing quirk of GHPC's crew AI in the past was that they could only become aware of things in the driver or gunner's vision arcs. We have plans to properly expand this to include the commander, but in the meantime, it was possible to simply drive up behind the enemy - even touching their vehicle - and watch them remain oblivious. That is no longer the case.

Crews now have a simulation of basic point-blank target detection, due to doing things like looking around through periscopes, or hearing a nearby vehicle on the move. At extremely close ranges, they will detect you instantly, unless you keep proper concealment between yourself and the enemy.

AIM TIME SIMULATION

Once a target is spotted, the crew still need to line up a weapon and perform some steps before a shot goes out. In the past, this was heavily abstracted, to the point that many tanks were performing complex manual rangefinding and gun laying calculations at a pace that would barely be possible even for a trained crew. (Many people have asked things like, "How are these T-55s shooting at me so rapidly from 2400 meters?") Chasing down and adjusting these settings was unnecessarily clumsy, so we created a new standard, which is both easier to tune and more realistic.

Each vehicle is now tagged to indicate what kind of systems it uses for rangefinding, lead, and elevation adjustment. The process is carried out in sequence, with sensible delays for each step. For example, the crew of the T-72M must go through this sequence: (1) Perform laser rangefinding; (2) Allow the range wheel and reticle to move to the desired range setting; (3) Perform manual lead compensation with the reticle.

An understandable reaction to this information might be, "But wait! There are plenty of cases where you don't need to do all of those steps. What about shortcuts?" And I have great news: we thought of that too! There are a whole set of exceptions and scaling bonuses in this aim time system for things like being within battlesight range (skipping the range setting exercise), engaging a slow or stationary target (reducing or skipping the lead process), and even engaging a target that's at a similar range to the previous one (skipping rangefinding entirely). We've set the AI behavior up so that the crews will act as quickly or methodically as they need to under the circumstances.

The result of all this intricacy is that every vehicle's rate of fire is now indicative of its tech level and features, applied in a consistent way that matches what you as the player can do when operating the same tank.

AI IMPROVEMENT IS ONGOING

Reading all of this might bring to mind related ideas for how the AI crews could perform more convincingly, or other dimensions of "realistic AI" that a vision and comms update doesn't address. Rest assured, we've thought about the issue extensively and have a list of other features to explore. What you see in this update is just the set of adjustments that we feel are ready to debut right now.

We're also monitoring the way people's gameplay experience changes as a result of all these adjustments. If issues arise, we'll address them as they come. The feedback channel on our Discord server and the general board on Steam are great places to give constructive input on how this update affects the game.

 

[cunumdrum]

22 hours ago, KillerBee said:

For the B29, depended on the variant and production date, anywhere between 10 and 20 .50cals with computer aiming assist and a 20mm canon stinger should be dropping some of those MiGs, but I didn’t “see” a single one firing.  These guns should be brisling with fire; in formation with other B29s, should be a wall of fire; more so than the B17.  The MiG tactic was to blow through the formation vertically, not come horizontal with it and take all that opposing fire.  And yes, explosions look very much 2-dimensional.

 The B-29's did not carry the 20mm in the tail. It as removed very early on in WWII as the trajectories were too different. I do agree I would like to see more firepower coming from the B-29's.

[LuftManu]

RE: Gunners

Guys, I think the mission gave not a good impression about the B-29 self defence capacities.

Jets are harder to track and shoot down with the defensive weapons (speed). And also, skill set matters.

This is what happens if we close in to a B-29 formation on veteran level, not even ace (On simple dogfight mode, they will be doing nothing) in a prop.
 

 

 

[Avimimus]

23 hours ago, KillerBee said:

For the B29, depended on the variant and production date, anywhere between 10 and 20 .50cals with computer aiming assist and a 20mm canon stinger should be dropping some of those MiGs, but I didn’t “see” a single one firing.  These guns should be brisling with fire; in formation with other B29s, should be a wall of fire; more so than the B17.  The MiG tactic was to blow through the formation vertically, not come horizontal with it and take all that opposing fire.  And yes, explosions look very much 2-dimensional.

I agree that they should be coming in from more of dive (as was the case in the accounts I read) and making only a single pass.

That said, the high closing speed would make the gunners less effective - especially for bombers other than the one directly targeted (as those gunners would have to lead the target)... and in effect, because the fire control was centralised, there are only one or two gunners aiming at a time... so there is the possibility of having a lot of guns miss.

That said, I think the gunner systems should probably be reworked eventually in order to better separate out estimation errors on the part of the gunner (lead, and especially ranging) and errors on the part of the gun system (e.g. difficulty aiming the gun at the estimated point, recoil effects). The system in Rise of Flight / Great Battles could do with being rethought. It'd even be great if AA guns could have the same types of ranging errors that were added for Tank Crew.

[AEthelraedUnraed]

8 hours ago, BraveSirRobin said:

When I say “good enough” I mean “acts human”.  The only area that I would expect any improvement is AI that has relatively human spotting/awareness.  But even that is probably extremely difficult.  Everything else will be obvious bot, probably forever.

Actually, I think that should be rather easy. Put a few people from the team in VR and monitor their spotting/awareness behaviour. I'm pretty sure it's easy enough to come up with a mathematical model that captures the intended behaviour believably enough.

Of course that leads to the followup question of what the intended behaviour should be. I'm sure people in a multiplayer dogfight environment have entirely different spotting behaviour than a P-51 pilot with his life on the line, halfway a 10 hour flight to Berlin. But that's a design choice rather than an AI modeling issue.

In one of their Dev Diaries, the developers commented on how the IL2 AI was designed with a multiplayer-like fast-action experience in mind. I think that design choice is at least part of where the current AI "issues" come from. If the AI doesn't spot you half the time until you've started firing (pretty realistic in real-life WW2 scenarios), you're going to have less of them exciting fast-paced dogfights.

[BraveSirRobin]

9 minutes ago, AEthelraedUnraed said:

Actually, I think that should be rather easy. Put a few people from the team in VR and monitor their spotting/awareness behaviour. I'm pretty sure it's easy enough to come up with a mathematical model that captures the intended behaviour believably enough.

If it was relatively easy they would have already done it.  

[AEthelraedUnraed]

Just now, BraveSirRobin said:

If it was relatively easy they would have already done it.  

That's a partial non sequitur. As far as I've understood, the AI system is being largely redone. It's absolutely possible there are earlier design choices that make it hard to implement a better spotting AI in the old, existing system. But that should no longer be a problem if you build the AI from the ground up. The fact that they haven't done a better spotting AI yet in GB, does not mean it is necessarily difficult to do.

Your prior argument that it's extremely hard to make AI behave believably human-like also doesn't hold. We already passed the point a decade ago where AI had human-like performance in very specific tasks. Recent research shows that humans are entirely incapable of distinguishing between human-made and generated images. For instance, in a 2023 study by Frank et al. less than 50% of all images were classified correctly: https://arxiv.org/abs/2312.05976

I can tell you that generating images is a much more complicated task than spotting.

[BraveSirRobin]

1 hour ago, AEthelraedUnraed said:

It's absolutely possible there are earlier design choices that make it hard to implement a better spotting AI in the old, existing system. But that should no longer be a problem if you build the AI from the ground up. The fact that they haven't done a better spotting AI yet in GB, does not mean it is necessarily difficult to do.

I’d love to hear more details about how a new AI system could make an extremely difficult problem significantly easier.