Impressions of post-2.008 AI

[LukeFF]

I played the Patrol over Kalach mission from Juri's Yak-1 campaign last night, and the combat with the Stukas and 109s felt very believable, with each plane picking out its own target.

[busdriver]

I agree Luke, the new air-to-air AI logic is a significant improvement. 

[SCG_Space_Ghost]

Thanks for the posts, guys. I hope I have some time to test and compare some of these observations.

[unlikely_spider]

Has anyone encountered enemy AI that was unaware of you? Most historical accounts I've read state that a large portion of kills were made on opponents that were snuck up on. However in the missions, the enemy seems to always have perfect situational awareness. Or maybe I'm just a bad pilot

[Gambit21]

Welcome to every WWII flight sim ever

[LLv24_Zami]

Now after couple of days testing, playing SP missions has been more fun than ever. Ganging is gone for good.

[216th_Jordan]

Has anyone encountered enemy AI that was unaware of you? Most historical accounts I've read state that a large portion of kills were made on opponents that were snuck up on. However in the missions, the enemy seems to always have perfect situational awareness. Or maybe I'm just a bad pilot

 

Yeah, thats still not possible sadly right now. Problem is that AI would need additional routines for checking / scanning the sky for enemies.

[1./KG4_Blackwolf]

It's SUPERB!

 

But my preferred aircraft all time, my Go-To fighter is still the 1st release of this IL.2 brand... the LagG3 :-)

Join the club! We have cookies! Thin mints too! 

[FlyingNutcase]

I played the Patrol over Kalach mission from Juri's Yak-1 campaign last night, and the combat with the Stukas and 109s felt very believable, with each plane picking out its own target.

 

...the combat ...felt very believable...

 

And in the end that's what we're after in AI realm. It's good to see we've taken a step forward.

[Scojo]

Yeah, thats still not possible sadly right now. Problem is that AI would need additional routines for checking / scanning the sky for enemies.

Not necessarily. They could do something as simple as a flat probability, all the way up to something as complex as you mentioned. There's a big spectrum and I think the devs could find some level of this feature to implement. And I think the effort would be well worth it since a feature like this is very impactful on the gameplay and simulation feel.

 

I think at the very least having a probability to spot based on relative distance, direction, and AI experience level could be a low cost option while still adding quite a bit to the gameplay

[Gambit21]

I believe you are correct Scojo, as obviously the AI already have logic in place to detect enemies - is a basic proximity trigger at this point.

The trick is to circumvent/modify that trigger/logic so that it only kicks in based on a probability.

[-SF-Disarray]

Wouldn't a timer work as well? The AI knowing a plane is there and reacting to that plane are two different things. Putting in a random delay on that reaction, maybe based on AI skill level, may produce a convincing lack of situational awareness.

[Albino]

Not necessarily. They could do something as simple as a flat probability, all the way up to something as complex as you mentioned. There's a big spectrum and I think the devs could find some level of this feature to implement. And I think the effort would be well worth it since a feature like this is very impactful on the gameplay and simulation feel.

 

I think at the very least having a probability to spot based on relative distance, direction, and AI experience level could be a low cost option while still adding quite a bit to the gameplay

 

Yes - this is an adequate low cost solution.

[Gambit21]

Don't be so sure it's low cost - it all costs money and what was layed out there by Scojo, and I mean all those variables is not a simple, low cost solution or we'd have it right now.

[sinned]

You got it Bearcat,

All good hearted comments but I see where you are coming from.

Spacyghost and I are good friends.

[TheSNAFU]

I flew 4 missions in PWCG and saw no ganging. In fact the combat was more believable than before. Really exciting fighting with wingmen that sometimes took it to the enemy and other times the other way around. In one mission I was badly shot up attacking Russian bombers and was able to leave the scene without half the Soviet Airforce chasing me all the way back to my base.

Very much improved!

 

Thanks to the Devs for confirming the problem and fixing it.

[Scojo]

Just did a Sturmovik mission. The AI felt good. When I was in the flight, enemy fighters attacked the leader. Once I separated from the group, they then targeted me, but only one or two of them. The other fighters picked targets.

 

My AI friendlies hit ground targets, but they still circle the airfield until they run out of fuel at mission end, though. This has always been a problem, even in RoF.

 

 

Don't be so sure it's low cost - it all costs money and what was layed out there by Scojo, and I mean all those variables is not a simple, low cost solution or we'd have it right now.

Compared to having search routines for the AI pilots, this absolutely is a low cost option. And just a flat probability to block the enemy in visual range event that is likely already in the game would be even easier to implement than what I originally suggested.

 

I'd encourage the developers to consider some kind of feature like this as it would add a lot to the experience. They can definitely find a level of implementation that is satisfactory while also providing a good ROI since there are a lot of people that play this sim's single player

[Lusekofte]

All the bombers do that in QMB, unfortunately. Which is why rolling the old campaign mission generator into the QMB would make sense.

 

Are you sure, I figured the same, until I remembered I had to choose a load out for them. Then I think equipped with bombs they behaved a bit more like bombers 

The IL 2 I suppose to escort in qmb behaved like aggressive fighters until I put bombs on them

Edited March 4, 2017 by 216th_LuseKofte

[Gambit21]

 Compared to having search routines for the AI pilots, this absolutely is a low cost option. And just a flat probability to block the enemy in visual range event that is likely already in the game would be even easier to implement than what I originally

"Compared to" being the operative term there.

It all costs money and I think Jason would be the first to tell you that if it were that easy and cheap, we'd have seen it in RoF already, nevermind BoS.

[Scojo]

"Compared to" being the operative term there.

It all costs money and I think Jason would be the first to tell you that if it were that easy and cheap, we'd have seen it in RoF already, nevermind BoS.

I SAID LOW COST USING LOW AS A RELATIVE TERM. I UNDERTAND THAT EVERYTHING HAS COST ASSOCIATED WITH IT. 

 

All I'm saying is they can find a low cost option and I think that it would add a lot of benefit to SP AND POSSIBLY BE WORTH THE COST, and I would like them to consider it

Edited March 5, 2017 by 71st_AH_Scojo

[busdriver]

Eureka...in the 10 days of Autumn Campaign I was able to get my wingman to strafe after he had killed an I-16. Very cool!

[curiousGamblerr]

I SAID LOW COST USING LOW AS A RELATIVE TERM. I UNDERTAND THAT EVERYTHING HAS COST ASSOCIATED WITH IT.

 

All I'm saying is they can find a low cost option and I think that it would add a lot of benefit to SP AND POSSIBLY BE WORTH THE COST, and I would like them to consider it

Your Sshadow impression is on point...

[Scojo]

Your Sshadow impression is on point...

Sorry I got frustrated...

 

Another thing to note is that hundreds of games already do this and in much more complexity. Plus most of those games aren't even simulations. I don't think asking them to consider the possibility of adding some form of this feature would be bad at all.

Edited March 5, 2017 by 71st_AH_Scojo

[Gambit21]

Not sure why you're getting frustrated - I comprehended your typing perfectly well and responded accordingly.

All I meant is that "low cost" is still significant. This isn't a new idea - they've been getting asked about this for years.

[Scojo]

Have they said anywhere that the simplest form of the feature is still too costly for how much it would add to the SP experience?

[Gambit21]

Jason has responded to AI inquiries on various occasions.

They no longer have a dedicated AI guy, and changes to AI like you describe are not practical at this juncture.

 

However they keep surprising us so who knows - we can hope.

I'd like to see it too.

[coconut]

Regarding the problem of AI excessive awareness, here's my input:

• A simple non-directional solution based on probability alone is no good. People tend to overestimate the benefits of random-based solutions. They think of the nice aspect of having the desired behaviour trigger when they want it, but not of the alternatives that will also be visible, namely that a) the "desired" behaviour triggers when it shouldn't, or b) the desired behaviour not triggering when "it should". In this case, a) would be "the AI did not see me even though I was flying in his 12", and b) would be what people are saying now "You can't take the AI by surprise"
• The directional solution is actually cheap. It's no biggie computing the dot product of the relative position of nearby aircraft with the "six o'clock" vector of other aircraft. One thing to consider is that the cost grows quadratically with the number of planes in close quarters, but I suppose that's exactly why we have the 10km visibility bubble. I still don't think it's a problem in SP since you have at most one player and 16 AIs, and it can be mitigated by spreading the computation over multiple cycles. There is no need to run the check 50 times per second, once every 2s or so should be enough.

Note however that by "cheap" in the second bullet, I mean "requires few CPU cycles". It still takes time to document, implement, test manually, implement automated tests to detect when the feature breaks after new development...

 

I work as a software developer myself, and I can notice how users get frustrated when very simple things don't get implemented. "It's so simple, and I suggested it 2 years ago!". The thing is, when your plan is full of big things to do, you don't have time to do the small things. It does not matter how small they are. Often these impatient users are themselves software developers, so it's not a problem with a lack of education. It's just that we tend to have less patience with others than we have with ourselves.

[Scojo]

I get all that. I'm not whining for more features, my point is that I think this is one AI improvement that would greatly improve the feel of the sim's SP and I think it would do so enough to warrant giving this a part of the dev cycle for implementation.

 

We're not talking about something quite like programming AI to estimate Energy in a dogfight, but rather just simply have something in their logic that will block the activating or receiving of an "enemy in visual range" event

 

As far as many planes being near each other, it would affect performance, however I feel you could deactivate the running of this feature once plane density gets high because in many cases the enemy's position would be relayed via radio, or the change in flight pattern of one or a few aircraft would warn other aircraft there are enemies in visual range.

 

If the devs really don't think it's viable at all or worth the effort, then that's fine. I would just like it to be considered.

 

As others have said, it's frustrating to have a perfect approach on an enemy, 6 o'clock low  or high in intermittent clouds, and the AI always knows you're there as soon as you're in their detection range

[curiousGamblerr]

 

 

my point is that I think this is one ... improvement that would greatly improve the feel of the sim ... and I think it would do so enough to warrant giving this a part of the dev cycle for implementation.

 

 

The thing is, just about everyone on this forum has their own favorite little idea to which they apply this statement. I'm not trying to be snarky, because a lot of these ideas are great, but coconut really hit the nail on the head. In a big project, the small things aren't small because of the testing overhead, and if you get caught up in all the small ideas you miss the big picture and end up with a mess a la CloD.

 

I'm looking forward to seeing what sort of things mod-on mode exposes. I'm doubtful of this particular improvement being possible, but I hopefully a lot of the small things people are pining for will be addressed by mods and end up with each of us getting closer to our ideal Il-2 experience. 

 

Anyway, don't get frustrated man, we know what you're saying. I was just having a little fun with my previous comment.

[Tuesday]

Going back quite a bit - I've never experienced the bombers behaving like fighters in QMB, only the new 111 H16 does that for me.  The rest continue to fly like bombers, as they always have. 

 

Discrepencies?  If anything it probably makes QMB more interesting. 

 

It would be nice to see even more improvements made to the AI if it were possible, absolutely.  I'm dreaming of course, but I like the player-summoned AI bombers that sometimes include a bit of an AI escort, too, found in some of the servers in CLOD. 

 

It won't make or break the game but it gave a little extra to my experience in CLOD: a pressing reason to scramble in an attempt to intercept and defend.  Whether giving something for a low-populated server or trying to divert enemy cover during a push in a different area. 

 

Hopefully (for me, at least) that is one of the things that can be added with mods-on that others might be interested in as well. 

Edited March 5, 2017 by II./JG77_Tuesday

[BlitzPig_EL]

Just ran a couple of the stock Moscow campaign missions.   Very much improved.

The AIs did their jobs on both sides, and the Bf 109s didn't all chase me!!

 

It was fun to fly offline, which for me is a pretty new thing.

 

Well done.

[unreasonable]

Are you sure, I figured the same, until I remembered I had to choose a load out for them. Then I think equipped with bombs they behaved a bit more like bombers 

The IL 2 I suppose to escort in qmb behaved like aggressive fighters until I put bombs on them

 

The last time I tried this it was with an He111 with maximum bomb load - seeing it try to fight in the vertical was quite a shock! Was a while ago now, so perhaps this has changed.

 

Just on the whole AI randomness in spotting issue discussed above: I do not know how much is in BoS but RoF already had an element of this and it could be modded a little too. 

 

In base RoF it was possible that one AI wingman from your formation might spot an enemy aircraft and leave formation to attack it while the others continued to follow you as leader. See my short video "Cheating AI Huns" for this. In addition, there was a mod that made it possible occasionally to surprise an enemy AI by breaking it's search pattern - in effect making the AI lose SA for a while. Both of these features added some nice variability to the AI behaviour.

 

On the whole I find current AI in BoS believable during my last SP campaign play through, although that may simply reflect that things tend to happen so fast and visibility is so restricted that I am simply not noticing any odd stuff! If Gambit's tests have now reduced the ganging effect even further so much the better. Small, incremental changes can make quite a big difference to the quality of game-play so I hope we can still get improvements on this basis while recognizing that Jason has said that a deep review of AI is not planned.    

[PatrickAWlson]

Per the directional solution: it has to be a little more sophisticated than a straight forward "can't see at 6:00".  Better pilots are more aware of the situation.  In real life they didn't fly completely straight in order to clear their six.  In combat a good pilot develops a mental map of the situation.  They are aware of somebody behind them even if they can't see them.

 

With that in mind I would tie the scope of the awareness to the experience of the pilot.  A simplistic solution would be to let better pilots "see" at a greater angle than lesser ones.  I would also give lesser pilots a chance not to see even if the enemy is visible.

 

Let me see if I can sum this up: the odds of a pilot being aware combine the pilots experience, the situation (combat vs non combat), the visibility of the pilot's plane, and the angle of the enemy plane to the pilot's plane.  

[Blutaar]

Just restirct the view like our non AI view is restricted, maybe also disable icons for the AI, they allways know who is enemy and who is not, never ever i witnessed AI chasing a friendly.

The thing is, the "mental map" you described is the same for every AI and they all are data linked to eachother it seems, they allways seem to know where everyone is, try a "do like me" command when you are 20km or more away from them, they will find you and join formation.

 

And before someone comes and tell me view is restircted, yeah maybe a bit but they have no nose, no wings and no tail and in my opinion they can track all targets at once.

I see that everytime i play when the AI pulls up after a chase and still perfectly follows me with his big nose directly covering my plane and copy my movement perfectly.

 

Didnt no one ever wondered about the long tracking high g maneuver bulletstreams the AI is able to send into your cockpit, thats only possible if you have no nose in front of you like IL2 46 with that guncross only view?

 

Just test it, try to make a high g maneuver like a break turn and keep gs high, if you got an enemy AI on your back he will perfectly aim at you, follow you and hit you with sometimes long consistent bursts, i cant do that, i just struggle to hit in that scenario because the plane vanishes under my nose if i trie to get a solution so its aim, pull and shoot, all misses or a few hits but never a long consistent burst.

 

Ah before i forget, try to sneak up from behind, at some point the AI starts to notice you and eventually break to the left or right right before you are in shooting range, its funny if thats an IL2 without rear gunner, if thats not eveidence enough i dont know, guess im dumb.

Edited March 6, 2017 by Ishtaru

[BlitzPig_EL]

I think the best way to understand the AI is to stop anthropomorphizing "them".

 

"They" are not separate entities. In fact "they" are not entities at all.  It is merely the game engine controlling one, or several, pilotless virtual aircraft.

And of course it knows where you are all the time, and when you are in a firing position at your gun convergence range, and it always will have a perfect firing solution on you.  

 

Why?

 

It's the game engine, the very same one that generates everything on our screens, that has all the physics calculations for everything that moves in the sim, including where your controls are the instant you move them. It knows every parameter of your position and speed, everything...

 

The hard part is programming it so it will make mistakes, and generally be as unobservant, and irrational as we are.

 

The AI are just ones and zeros in a program,  "They" don't think, "they" calculate...

[Thad]

I'm a one and I don't get along with zeros.

[unreasonable]

Per the directional solution: it has to be a little more sophisticated than a straight forward "can't see at 6:00".  Better pilots are more aware of the situation.  In real life they didn't fly completely straight in order to clear their six.  In combat a good pilot develops a mental map of the situation.  They are aware of somebody behind them even if they can't see them.

 

With that in mind I would tie the scope of the awareness to the experience of the pilot.  A simplistic solution would be to let better pilots "see" at a greater angle than lesser ones.  I would also give lesser pilots a chance not to see even if the enemy is visible.

 

Let me see if I can sum this up: the odds of a pilot being aware combine the pilots experience, the situation (combat vs non combat), the visibility of the pilot's plane, and the angle of the enemy plane to the pilot's plane.  

 

 

Some of this seems to be already be there in BoS - not every AI fighter you engage is jinking about, sometimes they appear not to have seen you at all. This may be more obvious when playing campaigns than in QMB. 

[Pail]

 

 

sometimes they appear not to have seen you at all.

 

Even after being shot several times... I guess they are outside the zone where they are told to engage or something.

[Feathered_IV]

That sort of thing is particularly noticeable in Rise of Flight. As soon as you enter firing range, you pass through the enemy's reaction bubble which triggers their "roll left and make diving turn" maneuver. If you stay on them you get a sort of merry-go-round effect all the way to ground level.

[PatrickAWlson]

The AI might be zeroes and ones but the trick is to make those zeroes and ones come to conclusions that are are humanly plausible.  A lack of perfection is the first part.  Randomization is a key element to imperfection.  Limitations are another key element.

 

A novice AI should almost always miss, bit sometimes hit.  You do this by expanding the "open fire " zone and randomizing the "open fire" point.  Novice AI will therefore tend to miss.  Ace AI will have a much smaller zone, much closer to the target, making it more likely to hit.

 

Novice AI should not fly to the edge of the envelope.  Ace AI should.  You handle this by limiting novice AI control inputs.

 

Novice AI should not have the same set of maneuvers available that Ace AI has.  You do this by restricting the decision tree based on AI level, such that Ace AI has the full set of options while lesser AI has increasingly limited sets.

 

Not easy by any means, but certainly possible.  If the decision tree and decision algorithm is already in place then it's not even that hard.  it is really precisely what chess programs do.  Pare down the decision tree before passing it to the decision algorithm.  You have immediately differentiated between AI capabilities and introduced a "human" element to the equation.