AI Discussion (dev branch)

[Tonci87 163]

Your report is valid. It shouldn't. After all - we don't have any vital-signs live-stream of subordinates ;)

Afaik it's generally this http://feedback.arma3.com/view.php?id=14905 and interconnected issues - things that reduce the enjoyability or give the impression of cheating AI. (nerfing first, pimpin' later ;) - that to say would be too simplified)

The UI work is more about adding the ability to properly test the AI, add standards and some reliability. Not locking down current AI, more about locking down some variables to allow for AI improvements.

In: Ezekiel 25:17 ;)

Armor simulation is a good example (and I believe it has been improved in A3). Others are sometimes better as "optional" or modular features, because they don't always fit everywhere. The community solutions generally (at least today) present a great inspiration to us and indicate what is missing in the vanilla game. Though it is still valid to ask "why we didn't" this discussion doesn't belong into this thread.

I really enjoy reading what you guys are currently doing and what your plans are, I wish there would be more of this information sharing.

Keep going!

Armor and penetration simulation has really improved a lot. The progress is very impressive.

You know what would be impressive too? Bipods and Weapon resting. Sorry couldn´t resist :p

But seriously why isn´t that in the game yet? Is it because of the AI? Feel free to drop me a PM

[Bouben 3]

Hi,

because AI drivers are still so problematic they are able to immobilize their vehicle by damaging their tires by constantly crashing into stuff. You can not take for granted they will make "those ordered 200 meters" without crashing their wheels (which leads to extremely slow movement or the crew disembarking from the vehicle). This is a crucial problem if you need them to behave reliably with their cars. Too often I lost my armoured car support because of this stupid AI behaviour.

I propose the following workaround, so that we can enjoy the gameplay, until some better solution is done:

Please, disable damaging of tires and wheels from crashing into stuff if the driver of the vehicle in question is an AI.

If this is possible and done properly we would still be able to damage AI's tires and wheels by gunfire and mines but AI would have significantly more chance (even after a lot of crashing) to bring their vehicle where ordered and we would not have to babysit that much every car ordered to move somewhere.

What do you think of this? Would it work for you? I don't have any better idea for now but I am feeling very frustrated that my AI mates are not able to keep their wheels alive even in peace times.

Thank you.

[.kju 3239]

Would be easy to script to see how it works in practice.

In addition I would advice using AI driving behavior in safe mode both for cars and tanks a lot more.

See my CIT tickets about this. Works a lot better for traveling jobs outside of combat situations.

[bravo409 13]

I have looked into that issue right after the first vid, the issue is wedded to the building. If you wan't to keep an eye on the progress, please start a FT issue and let me know via PM so I can tell you once it has been fixed.

In the latter videos you are dealing with completely different issues. What alarms the AI and what doesn't (based on audible sources). And how is the information shared across the groups.

But that is the most complex ;) I'm sorry to disappoint you, but human-like, even monkey-like behavior is still beyond reach even of the academic level AI. And that itself is couple of years, decades, ahead of any AI usable in an open game environment & with home-PC computing power :/ Many things are not (often can't be) simulated in Arma 3 on the same level how player perceives the game (e.g. not everything the player hears, the AI hears as well).

I'd like to ask you for your patience. We are constantly working on the AI, trying to improve it and your feedback is of a great value to us!

The most helpful is of course a isolated 100% repro about objectively (also present-day game-wise) wrong behavior, as simple as possible, 2-3 AIs... ;) submitted into Feedback Tracker . That is like a golden nugget to us ;)

But by far not only - a description of general experience helps us greatly to decide on the course of action we should take, what to prioritize, even if we can't eliminate the variables that can count towards it (subjective matter, skill settings, mission, mods, scripts used...)

---------- Post added at 14:30 ---------- Previous post was at 14:24 ----------

I can assure you we do :) Sadly not every scripted solution can easily make it into the engine as a full-scale game-wide mechanic :/

I'm sure the ai can be fixed to what people want to see just look at KAI. The once thought arma 2 that couldn't be changed ai wise is awesome with KAI ...unfortunately it's not for public they say.

[Bouben 3]

;2571289']Would be easy to script to see how it works in practice.

In addition I would advice using AI driving behavior in safe mode both for cars and tanks a lot more.

See my CIT tickets about this. Works a lot better for traveling jobs outside of combat situations.

So you volunteer for the script writing? :-)))

Please, post links to your tickets.

[froggyluv 2130]

I'm sure the ai can be fixed to what people want to see just look at KAI. The once thought arma 2 that couldn't be changed ai wise is awesome with KAI ...unfortunately it's not for public they say.

What the heck is KAI? Your saying a revolutionary AI mod exists that is not available to the public :confused:

Curious if the Dev's are working on eliminating extraneous actions by the AI -meaning, they were supposedly working on streamling the AI but not adding new features. Personally I would be ecstatic if by streamlining they meant more deliberate, decisive action and less leaving good cover and deciding to walk into the path of incoming fire.

[-Coulum- 35]

Unfortunately I don't think this is a "streamlining tweak" but rather a ton off extra coding. It seems that the ai leaves cover because they are meant to leave (in order to stay in formation.) not because they are poorly streamlined or have too much going on around them.

And unfortunately its not as simple as making formation looser or less priority because this will also introduce flaws (groups being spread by 100s of metres) and plenty of frustration when commanding the suckers.

Right now it seems like most effort is going towards ai skills. And its a tough task. A big step forward on that front would be to make the ai more accurate but slower firing with magnification and less accurate and faster firing without.

Was wondering what KAI was as well. Found it in ChrisB's sig.

[froggyluv 2130]

I appreciate your feedback and long term commitment to the AI Column without having soured :)

Also extend the same to the AI dev branch team -it's great that there's actually a group designated for this, the hardest most challenging part of any game where most have literally abandoned it. I guess what I'm not getting is that if formation and possible lost command, need for regrouping a scattered unit that has just successfully performed a firefight by utilising successful individual cover routines is the tradeoff -I'm MUCH prefer the latter. there is little point in the satisfaction of a well oiled formationed group that can't perform competantly in firefights. I see too often an AI have a great angle, using a wall for cover with lean outs against a prone enemy 40m out -the cover guy almost without fail will eventually just abandon cover and walk out into fire = prone guy wins everytime. Hard to build a tactial plan for that outcome.

These guys need to be unshackled from formation for firefights especially in Danger mode. I believe it's FabrizioT who has a ticket for allowing more liberal control of how to modify behavour in spite of formation when in Danger. I hate to be one of those "Arma 2 AI mods rule" guys but I've certainly seen Ai units hug to cover far more effectively in spite of squad formation in them. "Break Formation to Cover" has got to be implemented at some point. Let "Return to Formation" handle it :D

[.kju 3239]

Cars/trucks not sticking to roads in combat mode

Make Tanks drive on roads in Aware mode

Remove the tank driver's AI collision avoidance when given a move command by the tank commander

AI drivers fail to detect obstacles in their way

Tanks are slowed down too much and unrelated to the speed when hitting world objects

AI shouldn't avoid running animals over when at risk

Wheel vehicles are not controllable by your AI driver when you drive them as a commander on steep slopes

Wheeled APC commander maneuvering/steering through terrain difficult to impossible

Unfortunately too much else to do these days. :(

[-Coulum- 35]

I guess what I'm not getting is that if formation and possible lost command, need for regrouping a scattered unit that has just successfully performed a firefight by utilising successful individual cover routines is the tradeoff -I'm MUCH prefer the latter. there is little point in the satisfaction of a well oiled formationed group that can't perform competantly in firefights. I see too often an AI have a great angle, using a wall for cover with lean outs against a prone enemy 40m out -the cover guy almost without fail will eventually just abandon cover and walk out into fire = prone guy wins everytime. Hard to build a tactial plan for that outcome.

Yeah sorry I don't mean to shoot ideas down I actually agree. I was just pointing out that its probably not part of "eliminating extraneous actions by the AI -meaning, they were supposedly working on streamling" but rather the opposite - introducing more variables to tell the ai when it is okay and not okay to leave formation. decreasing streamlining when AI aren't following you in battle or won't heal that squad mate because he is stuck in cover.

I do wish formation vs. cover was improved upon. Even better I wish formation was based upon cover - Ie. a squad leader will decide to stop if under fire and in an area dense with cover so his team can take and hold cover till the fire decreases. And it doesn't just seem to be formation that screws them but also enemies - if they have an enemy in their sight they aren't moving(to cover) until they feel they have engaged the enemmy sufficiently. I think this is one of the reasons AI seem so poor at taking cover - how are they to take cover when every-time they see an enemy they refuse to budge but when they don't see an enemy they don't know how to take cover from it.

I think the ideal would be to make dynamic formation limits based on incoming fire. If a team is taking heavy fire then ai say screw formation and stick to cover. If things are going dandy they stay in formation. Unfortunately I extremely doubt this is simply a matter of tweaks though.

But it is one of those things that will need to be changed eventually. I agree with fabrizioT, that the ai behaviour and regard for formation should be more open to modders because without it mods are forced to use scripting commands to fight the ai's instincts to stay in formation. And though it works, its not exactly pretty.

[bravo409 13]

What the heck is KAI? Your saying a revolutionary AI mod exists that is not available to the public :confused:

https://www.youtube.com/channel/UClTOJWvzvTL6hTTO5S20vSw

BTW they made this for themselves so doubt we see this unless there are high demands they make it for the community ...Also dont hold your breath for on Arma 3 there more interested in A2 and said they may make it compatible with A3....

Edited December 6, 2013 by bravo409

[Bouben 3]

https://www.youtube.com/channel/UClTOJWvzvTL6hTTO5S20vSw

BTW they made this for themselves so doubt we see this unless there are high demands they make it for the community ...Also dont hold your breath for on Arma 3 there more interested in A2 and said they may make it compatible with A3....

I absolutely cannot see what is so great about it (KAI). The car in formation still goes first to catch all AT fire and infantry is not doing anything special too. The squad would sometimes be killed with one grenade. The infantry is advancing while their "covering" units are not covering at all (looking in wrong directions). AI ignoring corners etc. I think it is completely overrated, because there are the same stupid stuff that are happening in vanilla. Sorry.

Edited December 6, 2013 by Bouben

[maturin 12]

The squad would sometimes be killed with one grenade.

You apparently don't know much about CQB. The compound clearing is amazing (given that it's all procedural scripting) almost human.

But the reaction under contact in the open wasn't so amazing, from what I saw, which completely stands to reason. Looks like the mod might be best at setting up realistic firefights, but then the old ArmA vanilla kinks shine through.

[Bouben 3]

You apparently don't know much about CQB. The compound clearing is amazing (given that it's all procedural scripting) almost human.

I understand stacking up but that was just random packing of units together with no purpose. And amazing clearing? So ignoring corners is amazing? Man, this is just vanilla AI with different flavour and robotic moving through buildings. Doing "this" while ignoring "that" and vice versa. Nothing revolutionary and certainly nothing I would like to see in Arma 3. I don't want to come around as bashing the people behind KAI. I just don't understand the universal praise from the community. It is just another AI mod that does something on expense of something else. Interesting effort but as something that should be giving example for Arma devs? No thanks.

That said, until the AI is able to really use cover (not just stand behind it, but to actually peak and hide and return fire blindly etc.) there is not much of a revolution to be coming to Arma. Until then, I really don't care how fancy the movement of units looks like because in the end you still get those units killed in the open - because they cannot use their cover.

Edited December 6, 2013 by Bouben
grammar and stuff

[danil-ch 165]

AI uses road signs as cover.:( http://feedback.arma3.com/view.php?id=16449

[Bouben 3]

AI uses road signs as cover.:( http://feedback.arma3.com/view.php?id=16449

And also wire fences...

[danil-ch 165]

And also wire fences...

Can you make repro mission pls?:)

[Bouben 3]

Can you make repro mission pls?:)

Not now, sorry. Will try to do it later.

[[aps]gnat 28]

AI not controlling Fixed Wing Aircraft turrets

Once turrets and gunners are added to aircraft (Class Plane), AI are frozen, do not move the turret to track targets/rotate turret/move gun. Only if target passes in front of AI might AI fire weapons.

Turrets are functional, can be moved/rotated by players, can fire weapons.

Same bug as ArmA2 here;

https://dev-heaven.net/issues/29050

http://forums.bistudio.com/showthread.php?130929-Fixed-Wing-Multiple-Turrets-Testing-Bug

ArmA3 FEEDBACK TRACKER REPORT - Please Vote

Our B17's, AC130 and even modern laser painting etc of targets from jet fighters needs this fixed.

[fabrizio_t 58]

And also wire fences...

And gates.

It's like this since ArmA2 (or even earlier)

---------- Post added at 11:27 ---------- Previous post was at 11:05 ----------

That said, until the AI is able to really use cover (not just stand behind it, but to actually peak and hide and return fire blindly etc.) there is not much of a revolution to be coming to Arma. Until then, I really don't care how fancy the movement of units looks like because in the end you still get those units killed in the open - because they cannot use their cover.

Le me add my 2 cents.

Units are already able to peek around corners and i see them doing all the way, a bit too much, actually.

Problem is that when doing that they're locked in position and sticked into a pre-determined stance, mostly ignoring commands and incoming gunfire.

The fact units are killed in the open is not caused by them not being able to use cover, strictly speaking, but rather by artificial restrictions which are probably there by (bad) design, for instance:

1) if cover is scarce formation is not allowed to stretch enough to allow all units to reach some kind of cover.

Look around for gammadust findings about fsmFormation. I agree with him than fsmFormation should allow for dynamic distance thresholds.

Ticket: http://feedback.arma3.com/view.php?id=15297

2) whenever some enemy comes into sight, a unit is stopped for some seconds.

This is bad, since it makes unit completely oblivious of situation, not accounting for being in cover or not, having to advance or not, enemy distance and angle, formation spread, ... .

This is something i'm addressing in bCombat and i assure you that if you clear units to move more freely and some cover is nearby, they usually go for it, unless hardcoded targeting routines take over (here we have another issue, no control or overrride on currentCommand).

That said there are major problems with taking cover mechanics:

* Pathfinding to cover position is clunky and unreliable

* Buildings positions are not considered cover

* Many objects are wrongly considered as cover (poles, signs, fence, gates, bushes, ...)

* Editor placed objects are not considered cover

* In COMBAT mode targeting / attacking mechanics overlap with movement to cover, stopping it (you see units stop-and-go continually, before reaching cover)

Edited December 9, 2013 by fabrizio_T

[Bouben 3]

The fact units are killed in the open is not caused by them not being able to use cover, strictly speaking, but rather by artificial restrictions which are probably there by (bad) design, for instance:

1) if cover is scarce formation is not allowed to stretch enough to allow all units to reach some kind of cover.

Look around for gammadust findings about fsmFormation. I agree with him than fsmFormation should allow for dynamic distance thresholds.

Ticket: http://feedback.arma3.com/view.php?id=15297

2) whenever some enemy comes into sight, a unit is stopped for some seconds.

This is bad, since it makes unit completely oblivious of situation, not accounting for being in cover or not, having to advance or not, enemy distance and angle, formation spread, ... .

This is something i'm addressing in bCombat and i assure you that if you clear units to move more freely and some cover is nearby, they usually go for it, unless hardcoded targeting routines take over (here we have another issue, no control or overrride on currentCommand).

That said there are major problems with taking cover mechanics:

* Pathfinding to cover position is clunky and unreliable

* Buildings positions are not considered cover

* Many objects are wrongly considered as cover (poles, signs, fence, gates, bushes, ...)

* Editor placed objects are not considered cover

* In COMBAT mode targeting / attacking mechanics overlap with movement to cover, stopping it (you see units stop-and-go continually, before reaching cover)

Thank you for your "two cents". Nice info.

Units are already able to peek around corners and i see them doing all the way, a bit too much, actually.

Problem is that when doing that they're locked in position and sticked into a pre-determined stance, mostly ignoring commands and incoming gunfire.

Yes, this is exactly what I was talking about. If this is not solved (this would probably be a new feature...), then there can be only limited improvement in AI combat. Especially in CQB. This is an absolutely crucial AI behaviour. I mean, if you are static behind an obstacle then you can just go prone in the open with less weapon sway and lower profile.

Edited December 9, 2013 by Bouben

[progamer 14]

Nm, not relevant to the conversation.

Edited December 10, 2013 by ProGamer

[fabrizio_t 58]

(Pasting below a message posted into the wrong thread)

A rather big improvement would come from reworking the way AI handles its target.

From what i've experienced, AI usually follows this pattern:

1. spot enemy

2. identify as foe

...

3. stop when it's in line of sight

5. fire on it, eventually

6. move on

Point 3. is very weak though.

A unit should be stopped only depending on the outcome of many criteria, for instance:

* Being in cover or not;

* Presence of nearby cover;

* Being moving fast or slow;

* Distance from enemy;

* Distance from destination;

* Angle between direction of enemy and direction of destination;

Some practical examples:

* stop if in cover and enemy is relatively near (attack form advantageous position)

* don't stop if far from enemy and moving fast to some destination on its side (e.g. good for flanking)

* don't stop if moving fast and some cover is nearby (better to keep moving fast, but towards cover)

[.kju 3239]

Good point. Would be great if BI could give modders the ability to tweak that behavior via scripting to test, learn and figure out good ways to.

[Bouben 3]

Yes, very cool ideas.