AI accuracy: To good, to bad or good as it is?

[-Coulum- 35]

Fair enough. I have been thinking about that as well. It's nice to see the AI attempting to blind-fire or inspect the last known location but I've had the AI shoot me through a bunch of treetops, or crowns if you will, and I'm pretty sure none of them were looking at me.

Well it is very hard to be sure they never saw you or don't have an unobstructed LOS. remember, to the ai, just the tip of your hat is all they need to get your exact location. And they pretty much never forget where that location was or what your last movements were, so if you aren't unpredictable it will appear they are shooting right at you even through foliage. I am not trying to defend the ai, I am just very very confident that it was something other than the ai simply seeing straight through the tress, that is causing the problems. But of course the best way to prove me wrong is to show steps to reproduce!

Also, if the AI sees you and is alert, when you approach from a different angle there is a big chance that it will immediately start shooting at you. While a human is able to predict such things (i.e. the angles of attack) the AI seems to do it past the point of acceptable. I think.

Yep another pretty major problem. i'd say this is the closest you'll get to the ai cheating. Once they are aware of you they seem to have a sixth sense about your position. They could see you, loose you and you would sneak up on their six and, although it may not be immediate, they will "happen" to turn and look in your direction. There are lots of things wrong with the ai detection ability. Right now once you're seen, besides kiling the ai, there is no way to truly become hidden again besides waiting for a very very long time for the ai to forget about you.

And don't hide behind brown bushes, the ai can see through the very small holes in them, even from far distances.

[froggyluv 2136]

And don't hide behind brown bushes, the ai can see through the very small holes in them, even from far distances.

This is everyone's biggest fear -is my 'big pixel toe' visible? That is, can the AI see player with 100% certainty if 1 pixel is exposed but the rest of the body concealed? I know there is some sort of nearby clutter calculation to lower knowsabout but beyond that I just don't know.

[-Coulum- 35]

This is everyone's biggest fear -is my 'big pixel toe' visible? That is, can the AI see player with 100% certainty if 1 pixel is exposed but the rest of the body concealed? I know there is some sort of nearby clutter calculation to lower knowsabout but beyond that I just don't know

haha yep its all about making sure every single pixel is hiden from those suckers (although technically a toe wouldn't be enough for the ai - they need to see your head or torso - they don't recognize arms legs or guns). Above I am talking strictly about objects, not clutter. I haven't done much testing of clutter and ai, but the way I interpret it now is this:

You know how units sink into the ground when viewed from far away due to clutter. Well this is how the ai sees all the time. They don't see individual blades of grass obstructing you, they just see the ground rise up to cover you. If the ground covers you enough to make you totally invisible to them than they can't see you. Otherwise its the same rules as objects - if a single pixel of your head is seen above the ground they can and will get your exact location very quickly. I don't know if that makes sense but that's the way I interpret it. This is completely untested however with no real evidence to support it. Its just the feeling I get when playing against the ai.

[johncage 30]

I strongly disagree. Rather:

Make an acccuracy slider in addition to the skill slider, increase turn speed, increase detection time.

I say this because:

Making the ai overall less accurate takes away control from players, mission makers and modders alike - a mission maker might want a sniper that can headshot you at 800 metres. An addon maker might want to make a robotic soldier with super precise aim. a player might simply want more of a challenge. Simply decreasing accuracy across the board disallows all that. Instead, make it easier for players to adjust there ai accuracy via the difficulty options so everyone can be happy.

I don't think the reaction time of the ai so much a problem as their turn speed. I think you probably mean the same thing but reaction time would technically be totally dependent on your CPU's capabilities. It is turn rate that allows us to run circles around the ai, right infront of their nose.

I suggest detection time over detection range because it is not impossible for players to spot ai from over a click away, especially with optics. So the ai should be able to do this as well. The problem now is they do it far to fast. It might take me minutes to spot a moving ai at that range and many times more if he were stationary (assuming he hasn't fired). This should be the same for ai. Currently they seem to spot you very quickly even at far range after you have shot off.

the ai doesn't work like that

[-Coulum- 35]

the ai doesn't work like that

I'm listening... what have I described that is incorrect?

[Moroe 10]

Too accurate for me at this moment.

Yeah. I aggree on both.

Overall I think is good AI for the early stage of the alpha, but need to be tweaked.

I agree

[krycek 349]

Have you guys noticed that with every Arma release the AI has insane accuracy even on low levels before they start the tweaking?

I wonder why they don't implement what worked in A2 1.63 and work from that solid base(I'm referring here only about infantry AI) instead of copy-paste the AI but with same issues as A1&A2 1.0 versions.

For example they don't pinpoint your position where they last saw you and aprox where you might pop your head because they currently have the old terminator vision.Me as a squad leader with bird cam and moving along a wall I can clearly see them following me as I move.True on low skill it doesn't happen everytime,but it's not rare either.It's not like I'm constantly stepping on broken glass.:p

Not only that but try to stay with your AI teammates on a hill and put some enemies at the base of the hill.After enemies start to move around and my squad spot them I can see my men pointing their guns into the ground meaning they can track the enemy in the lower part of the hill.

[Rydygier 1317]

As for detection...

I think, that to achieve more human-like spotting (and consistently reaction reflex/firing/hitting) ability there should be considered some additional factors, that limit human player (do not know, perhaps these are considered somehow, but effect is, like these aren't).

1. Recognition

Crucial. I'm not saying, that all this should be reflected accuratelly in game, but somehow considered should be target's surrounding, background (colour, contrast, distinction - for avoiding that camo patterns are used after all), kind, size, movement, emitted sound (and audible sound background, that can conseal given sound. Sound of certain vehicle will be not an attractor, when there is many sources of similar sounds around), but also smell (bodies, smoke), that will not give any precision bonus directly (unless linked with wind direction), but will make observer more aware and looking actively for the source...

2. Subsequent objects recognition over constant and simultanous

And there should be a chance for bad recognition, eg considering hostile as an ally from the distance and not paying more attention on it until he will do something, what will compromize him or at least bring observer's attention on him again. If such knowledge collecting is in progress, starting same process for another, not close to the first, object simultanously should be far harder for the same observer (as he is focused on another object/place). Moreover - if some object will attract all attention, there will be easier to simply miss out another, especially not so evident. Fact, that knowsAbout should rise with its consequences as is curently only, when observer will determine given object as worthy of attention already reflects active efforts to cognize, what caught my attention. This brings us to the...

3. Attention

It is obviously not only about having not obstructed LOS. "What is this, what I'm looking at?" It is the question not asked all the time when I'm "scanning with my eyes surroundings" for sure. Firstly needed is calculation, if observer will recognize, that this, what he see, is an object worthy of attention at all. Attention must be triggered. Only then he should analyze, what exactly is it, if anything (yes, with chance to focus on something, that only looks suspicious, but in fact is nothing particular). This attention factor is very interesting here. If I'm concentrated on something else, than observing given area/spotting, it is needed much more to get my attention. Even, if technically I'm watching given direction/area, often I do this "blindly", without focus/awareness. If I'm under fire from unrecognized spot, I doubt, if I would be so often able to nail this spot until I find some cover and calm down a bit, enough to think and do some action for that purpose. In some situations I'll be able to determine general direction by what I hear, sometimes this can be unreliable too (echoes, panic etc). The more if this is about spotting someone else, than shooter. Similar, when I'm wounded, relaxed, pensive or simply bored (not watchful at the moment). In such circumstancies there is needed much more than visible part of object to make me awared about its existence, and even more to get my attention so I'll start "knowsAbout" procedure. Do unit/group behaviour (safe, aware, danger etc) makes any difference in spotting ability? Easier for the attention should be, when observer knows much about object kind, so he knows, what can be expected, what to look for...

4. Typicality

Far harder to determine interesting/suspicious unit in the crowd of similar units, also movement will not attract observer's attention, when there is plenty of objects in move there - crowd again, also unit in the forest during windy day, when all leaves/scrub are in motion; object placed in not typical for it place will surpirse, so will easier attract attention. This affects both, attention and recognition.

Too complex? Too sophisticated? Not for me, not, when this is about really remarkable and unique AI, not, if we want true human-like spotting abilities. Still, even some simplificated emulation of above factors would be great improvement. I'm affraid, that if this needs deep revolution in relevant code - then will probably not happen though. Well, everyone has the right to dream. Symptoms like described exact following with aim through the ground movements of the enemy saw by another team member lead me to the conclusion, that any of above can indeed mean full AI reworking, so...

But what, if such reworking sooner or later is unavoidable? If current AI stuck in dead end and simply can't be improved much more except some cosmetics and tweaks due to some fundamental limitations? Only devs know that...

Edited March 20, 2013 by Rydygier

[maturin 12]

This is everyone's biggest fear -is my 'big pixel toe' visible? That is, can the AI see player with 100% certainty if 1 pixel is exposed but the rest of the body concealed? I know there is some sort of nearby clutter calculation to lower knowsabout but beyond that I just don't know.

This is a big fat myth in ArmA 2. AI can't even recognize soldiers with everything below the waist exposed. Same goes with arms and gun.

So please don't be lazy and go test it first.

[-Coulum- 35]

This is a big fat myth in ArmA 2. AI can't even recognize soldiers with everything below the waist exposed. Same goes with arms and gun.

So please don't be lazy and go test it first.

Yes you're right about toes and arms giving away position - but that doesn't mean that a single pixel doesn't give away your position. From my testing, if you are hiding behind a bush and expose even the slightest bit of your head, the ai can and will see you even at long ranges. This is a problem because in the case of a bush, most people wouldn't be able to distinguish the very small piece of helmet from the bush itself, especially if it is at long range. Basically partial concealment and camouflage don't phase the ai.

[froggyluv 2136]

Lazy? First off the 'big pixel toe' was a joke so lighten up Francis...:rolleyes:

To make a point on whether if one pixel is showing -even if that be parts of the body that register ie..head or torso -does the AI see that pixel as you in your entirety or is it minimized somehow in their calculations. If you know an easy surefire test for this -please, be my guest.

[NeMeSiS 11]

Lazy? First off the 'big pixel toe' was a joke so lighten up Francis...:rolleyes:

To make a point on whether if one pixel is showing -even if that be parts of the body that register ie..head or torso -does the AI see that pixel as you in your entirety or is it minimized somehow in their calculations. If you know an easy surefire test for this -please, be my guest.

'Showing' is of course hard to tell, considering

Grass...

[froggyluv 2136]

Those are good video explanations and I've seen them before -mostly I'm just curious how the LOS raytrace works if that one or small group of pixels is seen thru a small gap in a viewblock. Would that intiate a full knowsabout or perhaps a question mark that 'something is there' worth investigating. I find this sort of technology fascinating :)

@rydygier: good post and agree with your assessment that that sort of information gathering/analysis would better improve AI believability.

Edited March 21, 2013 by froggyluv

[maturin 12]

That's just not that much of a problem, though. Because bushes aren't opaque. They are complex 3D objects and almost always partially transparent. The AI just gets a big solid block of blindness that usually favors the player in terms of volume (with bushes. Trees are the other way around in Chernarus).

For every hypothetical 'one-pixel reveal,' we will have killed hundreds of AI after seeing their feet or barrel, or just looking through the damn leaves. Logic is the best weapon. Of course, us winning is natural gameplay, while a single player death is always CHEATING DAMN SHITTY GLITCH GAME RAGE.