Jump to content
Sign in to follow this  
Wiki

AI Improvement

Recommended Posts

What is that showing? That when the AI are seen out of cover, there is no suppressive fire, no movement out of there and if these are too much to ask, the soldier on the right is still just oblivious to what's going on and does nothing.

Yeah, AI has come on leaps and bounds. I can see where all that processing power is going.

Share this post


Link to post
Share on other sites

Yes the AI are a lot more deadly now,mostly for the better but TBH when they call out "contact 2.5 klicks north" when I am running through woods and the AI they have spotted are also in some woods make it a tad unrealistic IMHO.

Mick.

Share this post


Link to post
Share on other sites
ArmA3 AI does not react to enemies being lit by lightsources at night

http://feedback.arma3.com/view.php?id=5048

Please upvote. Support adequate reactions for AI, support Thief-like stealth :rolleyes:

Voted! Now we can have tacticools action hero thiefs (like thief 4) ;-)

Share this post


Link to post
Share on other sites
What is that showing?

kyang03 claimed that he tested with 2 squads and they moved around a lot, shot a couple times and walked in the open and took about 10 minutes to finally finish a search and destroy mission in a open field. I believe my footage clearly demonstrates that he needs to adjust the AI skill settings...

Share this post


Link to post
Share on other sites

Is it really a fair comparison? I am not sure that A3 AI on A2 map would react (use cover, etc) the same as A3 AI on A3 map. Because of cell sizes, object names, building positions (for cover) et cetera. Not discounting all your efforts at testing, just reminding.

Share this post


Link to post
Share on other sites

So far the AI is better but still lacking

the other day me and some other people were trying to take control of a small town and we had an Ifrit and I was the gunner all the sudden 3 troops just ran right in front of my vehicle completely exposed and vulnerable. Of course like every ArmA player would do, I opened fire at them and killed them all...

My point is, No one in real life would have ever done such a silly thing if they had half a brain. I mean seriously unless the AI is scripted to suciding once it knows its gonna be defeated I don't know what was going on.

Please do something about it its weird and ruins the immersion.

Share this post


Link to post
Share on other sites
If BIS looked back at some decent ai mod’s they could get an answer to many of the problems stock ai have.

If they stuck GL3 or 4 in this game, they would have decent ai behaviour, at the moment we have nice movements, or not bad, but little content regards ai behaviour..

AI doing just a few bits of what they can do, i.e. flanking, groups communicating with one another, suppression, using buildings, engaging from windows, doorways, balconies, rooftops, finding cover, calling reinforcements, air support etc, etc. Its all been here for years in the series, but in the form of mod’s, BIS have never cracked decent ai behaviour for the stock ai.

Problem we have, is the makers of the best ai mod’s, have left and moved on..:(

Lots of it is below in the test vids 'Intelligent ai'.. plus 'Skirmish' etc... We will never see any of that in stock ai.. or for that matter in A3 in general I fear..

Have you ever looked at the scripts in GL4? there is a big difference between scripts like this that set waypoints etc, and behaviour fsms. Don't get me wrong, I use elements of GL4 too but part of the point of that is that it is configurable by the user. For me the issue is improving the ai fsms, I've no objection to continuing to use script frameworks like GL4 to provide dynamic group responses.

Share this post


Link to post
Share on other sites

Yes... i think it is important to keep different levels of ai clearly separated:

- individual behaviour and self preservation (lowest level)

- tactical behaviour adds to and influences the previous and determines at a more group level (medium)

- strategic behaviour - working towards goals (high level)

keep also in mind that the ai decisions at these different levels should also be representing different levels of human rationality/emotionality. What takes precedence over another in conflicting results has not a set answer.

as i see it...

- the 1st should be absolutely provided by the a3 engine, while it shouldn't be oblivous of what is happening at the tactical level, by default should take precedence. Some modability should still be provided here. (FSMs/whatever to the effect | SLX type ais)

- the 2nd should be 100% moddable but BI should provide its default of how it should work. (Scripting level | GL)

- the 3rd should be absolutely moddable yet BI could provide its vision (Scripting level | HAC)

I think this is basicaly what we actually have, the problem as i see it is that some higher level stuff is being defined at the low engine level (even if by ai necessity) yet lacking moddability, resulting in unxpected/absurd results in some cenarios which only hardly can be tackled. Clearly separating these levels at the a3 engine level is, imo, the first step to overcome the many issues.

Edited by gammadust

Share this post


Link to post
Share on other sites
Yes... i think it is important to keep different levels of ai clearly separated:

- individual behaviour and self preservation (lowest level)

- tactical behaviour adds to and influences the previous and determines at a more group level (medium)

- strategic behaviour - working towards goals (high level)

keep also in mind that the ai decisions at these different levels should also be representing different levels of human rationality/emotionality. What takes precedence over another in conflicting results has not a set answer.

as i see it...

- the 1st should be absolutely provided by the a3 engine, while it shouldn't be oblivous of what is happening at the tactical level, by default should take precedence. Some modability should still be provided here. (FSMs/whatever to the effect | SLX type ais)

- the 2nd should be 100% moddable but BI should provide its default of how it should work. (Scripting level | GL)

- the 3rd should be absolutely moddable yet BI could provide its vision (Scripting level | HAC)

I think this is basicaly what we actually have, the problem as i see it is that some higher level stuff is being defined at the low engine level (even if by ai necessity) yet lacking moddability, resulting in unxpected/absurd results in some cenarios which only hardly can be tackled. Clearly separating these levels at the a3 engine level is, imo, the first step to overcome the many issues.

Good post, I agree.

Share this post


Link to post
Share on other sites

The problem is that modders have done a lot of stuff that works great for AI and yet BIS is still extremely slow at implementing what's already there into vanilla.

So they've (relatively) taught AI how to kneel to fire and how to throw smoke when wounded? (something that we had in mods for years)

Yet they still haven't taught AI to share info between squads if those squads have soldiers with radios - something that makes playing ArmA2 with ASR AI a must.

You can have 2 squads fighting it out literally 50m away from another squad. And if there is a wall in between (can easily be tested at an airfield) that remaining squad will just stand there like nothing is happening.

I will do a repro of this if there isn't one but man... BIS - why so slow? It's already there through mods.

Share this post


Link to post
Share on other sites

BI requirements should in theory be more demanding, hence the slower developement (i am intentionaly excluding other reasons, for now).

When a modder implements some kind of AI functionality, in fact describing a decision process through scripting, one as a very specific intention to achieve (ie. throw smoke to conceal movement), and in most cases gets easy to describe its conditionality (what cenarios is this decision acceptable). But then a lack of sufficientlty describing the latter may result in odd behaviour (ie. for all the other infinite cenarios), but as players it comes easyer to accept it, since you're installing the mod to respond to the predicted cenarios.

Now BI must be much more constrained in implementing AI because the risk of unintended consequences is more permanent. To BI an holistic approach is constantly required, constant awareness of the whole (from a programming pov - what are 90% of the cenarios this programmed response will trigger, and can we sufficiently describe them all?).

So i totally accept the "slow" implementation, it is warranted. Failing this we would have much worse AI for doing smart things at totaly inappropriate times. That is why the ai implementation must be layered, from general decisions through the specific ones.

The other not mentioned reasons could be unaccounted/lagacy code, which is like an explanation for everything, still an explanation.

EDIT

I think it might be worth to add the following when i said:

"- the 1st should be absolutely provided by the a3 engine, while it shouldn't be oblivous of what is happening at the tactical level, by default should take precedence. Some modability should still be provided here. (FSMs/whatever to the effect | SLX type ais)"

Imagine i am a modder looking to improve the ai behaviour from scratch. Can i? Well i can disableAI, disableFSM and all that, but then the instructions for AI to do very basic level stuff have to be implemented from scratch too. This in a way reveals how meshed together are different levels of AI, when it would get easier to address AI if they weren't. One should not have to keep in mind the bullets flying by when implementing a new pathfinding system, since that should be already there in some layers below. At this level i could be worried with what is the general direction of the bullets, not if they are close your head or not. The example could be developed further but i think the gists is there.

basically if one could set the AI to:

- catatonic state (reacts to nothing)

- stupid state (reacts only by instinct)

- comprehends a language and stricly follows directions state

- able to take tatical decisions state

- able to devise plans to achieve a goal

- increasing layers of awareness/intelligence

At the level you pick any of these the next layer of implementation has to be adapted.

Edited by gammadust

Share this post


Link to post
Share on other sites
kyang03 claimed that he tested with 2 squads and they moved around a lot, shot a couple times and walked in the open and took about 10 minutes to finally finish a search and destroy mission in a open field. I believe my footage clearly demonstrates that he needs to adjust the AI skill settings...

Im not new to this and yes they are all the way up, as i said ASR AI would still be better

Share this post


Link to post
Share on other sites
Is it really a fair comparison? I am not sure that A3 AI on A2 map would react (use cover, etc) the same as A3 AI on A3 map. Because of cell sizes, object names, building positions (for cover) et cetera. Not discounting all your efforts at testing, just reminding.

You won't believe it but they will. I posted a video a few posts back about AI comparison and A3's AI on Chernarus do use all cover as possible.

Share this post


Link to post
Share on other sites
Have you ever looked at the scripts in GL4? there is a big difference between scripts like this that set waypoints etc, and behaviour fsms. Don't get me wrong, I use elements of GL4 too but part of the point of that is that it is configurable by the user. For me the issue is improving the ai fsms, I've no objection to continuing to use script frameworks like GL4 to provide dynamic group responses.

A mix of mods, not just GL4, but SLX, Zues, Tpwc will produce decent behaviour. Here’s a test done a while back, just two factions set to run at each others starting point. Would be a senseless little scenario if it wern't for the mix of mods included. Cover is provided in the form of two small settlements, set over to the side of each group 40+ mtrs away from the main run area that the ai will take, no other cover is on the map. What would happen in every scenario without this mix of mods. They would see each other, then at best, drop to the ground and move forward using WW1 tactics, shoot, raise a little, move forward, shoot, repeat until dead, thats not what I or my group wants to see, its pointless behaviour, could do that behaviour on a board with dice (infact used to a few decades ago). This is an extreme test, no options to hide behaviour, thats the point, we want to see the behaviour, how the parts of the mods interact.

Improvements have been made since this test to our groups ai behaviour. But we're in A3 now, these mods are lost (well perhaps), so we have to want/hope for better. Its new ai we have here in A3, buts its not really, its ai from a long time ago, just made more accurate to hide bad behaviour, the ai behaviour our group had in A2 was about as good as we were going to find. But if we are to move onto A3, well, what to do..

Anyway;

The test here shows some of what these mods can do when mixed. Gl4 is for cohesive group work, backup, flanking, well, too much to list here. SLX adds a little in terms of awareness, some decision making, stances, movement awareness, plus other behaviour. Zeus adds more ai skill (thinking), self preservation etc, tpwc helps with some of the suppression, plus stances. Combine them, or parts of them, to get a reasonably sensible ai..

Work your configs well to oblige each of the mod parts, then make missions that are really put together well. Do all that and you have a game completely different to A2 and now A3, in A2 we could change that in A3 we're stuck, at present..

But if BIS put anyone of these into A3 in singluar form, ai would improve, if they mixed them, the same as our group has, then I would be really happy..

The old argument is, well it breaks missions (GL4), well yes, but it doesn't break missions made with the format of GL4. Our group missions are far beyond what you see in the standard Arma game-play, there not huge missions, campaigns are broken down into parts, but they feel, look and are, far more realistic to game-play, than anything I've seen anywhere concerning the series..

Must add I don't play missions in third person, just for tests. First person all the way for realism..

Edited by ChrisB

Share this post


Link to post
Share on other sites
(BTW, why do we say "vanilla" version?)

When ice cream was invented it only came in one flavour for donkeys years. Then someone invented other flavours.

Now we apply "vanilla version" to subjects outside of ice cream.

Share this post


Link to post
Share on other sites

Say, has anyone actually tried utilizing the existing AI mods in the alpha just to see what would happen?

Share this post


Link to post
Share on other sites
Say, has anyone actually tried utilizing the existing AI mods in the alpha just to see what would happen?

Our group is working on that now, could be problematic here and there, especially with the older ai mods..

Share this post


Link to post
Share on other sites
A mix of mods, not just GL4, but SLX, Zues, Tpwc will produce decent behaviour
yes i get all that, I've been using them for years. what I'm saying is, have you looked at the scripts to see how it is achieved, say in GL4? If you do you will see how far removed it is from the behaviour fsms that are used for low level AI. It is the low level AI that needs work in some areas.

Share this post


Link to post
Share on other sites

For now, I'm happy if the AI knows how to stance adjust themselves... Right now they're still at ARMA2 level where the best thing they can do is leaning. (Not generalizing here)

Share this post


Link to post
Share on other sites

ok here are my observations so far.

what i tested was AI reaction to fire and the following tactics to "resolve" the situation.

first thing to notice. when you place only one soldier in the editor all he will do is turn to your direction go prone and fire at you if you are still in the open after your first shots. he will hold the position proning and do nothing but scan your general direction.

if on the other hand you place 2 grouped units or more (3 is best for testing) you will notice that the group leader will do the same as a single unit and hold his position prone while the other two advance in your direction. what i concluded from this is that actual movement is based on move commands issued by the group leader and not on real micro AI.

now when you, like me in i one of my tests, place yourself near a house on the airfield and them on the runway you will notice the following. after you shot at them and went to a hiding place they will, after cooling down from their first prone reaction, go after you (2 of 3). in the open, runway in this case, they will perform a silly cover and move routine which means one will cover from the prone position while the other will advance (only!) up to 4m. then they switch and repeat. this would at least look more natural if they would advance more like 10m or more in each cycle. i am aware of certain military tactics but in the videos included in this post you will see what i mean by silly.

once they reach the first potential cover they will suddenly start to move much more natural and that is the only real difference i can notice to arma 2. they have improved path finding in close quarters. while this looks almost human at times you will quickly notice that it is just cosmetic and doesn't really show a real sense of awareness or movement based on most certain position of the enemy. while they move well to flankyou, you will soon see that they aren't really aware of things when giving them new input in form of one or several gunshots. while they kind of seem to register the shot they won't really react to it and they won't be able to pinpoint your location better based on it.

in my tests i fired several shots while hiding inside a house when they were near. they will not search houses. while on very rare occasions they will go through a house or go to one of its positions and then leave it after a short time, they will never check all positions and certainly not look for you in the house you are most likely in nor do they seem to be aware of you being in there.

so now here are some simple videos showing what i'm describing before i come to my final conclusion...

this shows how the AI still struggles with close encounters even if they happen in the open. why this is still basically like in arma 1 is beyond me...

this shows the squad routine i described in transition from open ground to close quarters. keep in mind that this is sped up. this takes much longer on normal speed.

this is the same method applied in all urban environment. while the movement itself seems natural the reasoning behind it does not seem simple or predictable but outright random and nonsensical at times.

conclusion:

while the movement in urban environment seems to suggest an improved micro AI it looks more like single units are still not really aware but rather commanded by one single "mind". the way AI reacts to close encounters of the sudden kind ranges from 0 reaction to slow turning to face you to super fast snap on your head aim bot shooting. corners of houses are used efficiently but houses themselves aren't really seen as such and the very rare entering of them seems forced and fake since it can also result in a unit staring at a wall from 50cm in the prone position.

it almost seems that the single unit layer is just totally random while the squad leader gives moves commands based on the terrain. while detection based on line of sight works ok most of the time it still needs a lot of work. sound detection is very basic and does just seem to trigger some random routines to move around the general area of the sound source. it never seems like a single AI has a real sense of direction but is rather a drone send to a place by the leader and has itself no idea why and then reacts weirdly when confronted. if arma was an RTS this would be great AI if still very random but for a game like this it's just not enough.

seeing what is possible by using scripts (GL4 etc) on top of the existing arma 2 AI and comparing it to this makes me ask myself if BI is just missing the resources to really work on this in depth. and i also ask myself why they don't provide us with a less random and inefficient simpler AI we can build on better. it just never seems really convincing or reasonable what they do. a lot of times it just seems broken which can't be bypassed by any amount of modding. we need something much better or what we have but severely cleaned up and debugged.

having looked at the AI in detail i don't know if i should be optimistic because this is just an alpha and the urban pathfinding has improved a lot already or if i should be worried since this suffers from the same issues like already the ofp AI did (at least there we knew what we get though). all we need is:

- real convincing detection routines for sound

- less errors and randomness

- vanilla house search routine

- first contact routine (get to cover FAST when not under immediate fire)

. close range reflexes

i've given up on real indoors pathfinding but those things shouldn't be too much to ask.

anyways it's late. i tried to keep it short and failed. sorry guys. i hope some people will read this or parts of it and share their thoughts on it. i will provide more videos soon.

Edited by Bad Benson

Share this post


Link to post
Share on other sites

I imagine that this was performed using vanilla AI (please list the difficulty settings) without importing of AI mods?

Share this post


Link to post
Share on other sites

Can you make a template next time before the wall of text Benson? Simply put: What difficult? What mods? Name the variables, name the test conducted and then give a description and conclusion. You got parts of it right but damn that hurts my eyes.

Share this post


Link to post
Share on other sites

Rye is right regards the wall of text, but understand that when you get going, trying to list whats going on, you can get carried away.

Good tests though, just highlights how little ai have come on, they really did leave it to the mod makers to sort out. Most of the ones I would like to see in A3 are made by people who have left the community I think. So there lies our problem as a group, ai in A3 at the moment are really bad, its going to be like that for a while I think. I have had ai go and search, well check, buildings in A3 whilst just standing around. Problem is running it on from that. I have yet to try more tests, still making missions for A2, for the group.

A3 in its present form is unplayable for our group, would just be annoying, but we can hope. A couple of the group are looking at the mod’s we used in our mixes for A2, they are pretty good when it comes to that sort of thing.

Arma 3 is good for looking around, for really good missions though with believable ai, well, may just end up a tourist resort...:butbut:

Share this post


Link to post
Share on other sites

When testing AI, keep in mind that the AI reactions or planning depends on lots of factors.

For example, a weapon fire audibility set too low can cause the AI to be unable to guess the direction the fire is comming from, or if set too high, it can cause the AI to be able to pinpoint the exact spot even from very far away.

Similarly, incorrect visiblity/audibility values for animations can cause the AI to not being able to see the unit, or being able to see even when it "shouldn't".

The fact that AI is having trouble engaging over small distances, could be caused by improperly configured weapon or optics attachment.

Movement issues can be caused by map/terrain errors, like by incorrectly placed objects, preventing AI from getting to a particular place, or forcing the AI to take strange paths around, etc.

It's rather complex with lots of factors working simultaneously.

When testing AI,try same scenarios on different locations, try with different weapons/ammo (for both sides), different units, etc.

Share this post


Link to post
Share on other sites

100% skill for friendly and enemy in the difficulty settings.

i usually tweak spottime and things like accuracy in my missions but this is vanilla 100% no super AI. if you really think that those settings make a real difference when it comes to actual intelligence then you are fooling yourselves. we need to discuss actual flaws/bugs and needed features instead of trying to fix what is broken by tweaking some settings that won't help anyways. for those complaining about too much text i added moving pictures to look at and several short points of what is needed thus missing at the end. if that isn't enough i don't know. no one is forced to read the elaboration and opinion part. don't be so whiny please:p

Edit: yea they sure can shoot you with one shot in one second over huge distances in open fields. i don't see how that is a prove of intelligence though. put them in an environment where they actually have to think if you want to see if they are able to. i've been making detailed tests with AI since ofp. i really know how they act in any situation. those videos show the best outcome. lots of times they act just utterly stupid. i don't see how finding excuses is helping the discussion. all the possible causes you describe could have been resolved since ofp. people seem to forget that the engine already knows where you are. all that is needed is realistic conditions that hinder that info from reaching the AI. what we have is broken. it looks like something that doesn't work as intended...

Edited by Bad Benson

Share this post


Link to post
Share on other sites

Please sign in to comment

You will be able to leave a comment after signing in



Sign In Now
Sign in to follow this  

×