AI Discussion (dev branch)

[Enricksolt 12]

For me the best way to fix the "terminator" AI, the random headshots and the way too short fights is to make the AI more arcade. What i mean is:

- Remove the recoil for AI (Since the main problem is always the first shot, who is the most accurate, recoil is pretty much useless)

- Increase the AI's gun's spread (In Arma 3, as it is in Arma 2, the weapons shoot pretty much straight and that's the main problem i think)

- Don't make the AI shoot to kill (AI in Arma 3 almost never fire for suppress and that's the one of the main reasons the firefight are this short)

- Don't make the AI aim always for the head (Even at 300m+ the AI will always aim for the head, that's fine in CQB but not in open field battle)

- The AI must wait a few moments before shoot after they got a target lined with their ironsight (Too many times you see AI start shooting when they don't even have you lined with their gun yet, that's odd)

- Limit the stance of the AI in combat to only standing and crouching (AI should never go prone, unless they have to ingage a far target or take cover)

Most of these things (like the shooting delay and spread) should be based on the difficulty.

In vanilla Arma 3 you sometime see good AI behaviour, sometimes you don't (sometime they act like humans, others they act like complete dumbs and other times they become some kind of killing machine), this make the game seem unpolished and IMO the only way to fix this is to force the AI to follow a standard\default shooting behaviour.

[dnk 13]

AI just need to be set properly by mission makers. 0.5 accuracy is way too high for normal GI (and many servers have 0.7 as the minimum!). The correct range for non-snipers should be 0.1-0.25 for "aiminaccuracy" and "aimingshake". 0.25-0.35 for "elite" units, and 0.35-0.75 for "snipers" or other one-shot types. 0.05-0.1 is appropriate for non-military shooters (gangbangers or whatever).

Still, it's silly that a "super shooter" skill level is 1/3rd of the total. The devs really should consider adjusting it so that 1.00 is the new 0.35 and everything else scales up accordingly. Yes, this would make certain sniper units unrealistically poor shots, but that's such a minor issue compared to having every unit being a super shot.

[Bouben 3]

I have noticed that AI have this annoying delay between standing up, running somewhere and getting prone/crouch. It always takes them a few moments before they finally get down again. I don't know why it is happening but it should be fixed somehow. They should be hitting dirt so fast in combat not thinking at all. They can think later after they are prone.

Also, they are still standing too much in a combat. It is really annoying to see them as they stand up absolutely unnecessarily when in long distance fire-fight. I mean if they need to check (and I don't know if they are capable of it in the system) somebody behind terrain profile/obstacle they should do so only if not immediately visible target is in front of them that they can combat first. Only after that they should check for another ones (or some kind of QUICK peaking system should be implemented - ideally peaking from behind a cover but that is probably something we won't see in A3 maybe not even in A4, sadly).

Another really needed tweak is that after some patch the AI is preferring shooting to taking cover. This is a great tweak in close combat but it should not apply in long range combat. In long range combat they should get maximum possible cover first and only then to start shooting because I have seen a lot of situations when this unfortunate preference killed them.

They also need to stop thinking while taking cover. I don't know if it is possible but it should take them literally like one second to decide to change direction of movement when ambushed, getting prone and meanwhile with maximum speed getting behind a hard cover. I don't know why everything has to take AI so long. If it has to do with performance I understand that but it is so annoying and so bad looking and it is pissing me off and preventing from frequent playing since A2. Somehow in OFP I didn't mind their reactions while playing the whole campaign two years ago.

Oh yeah, I almost forgot this one: snipers are useless because they fight like regular soldiers. They keep running around, STANDING, making too many stupid things and are generally speaking very ineffective. They also don't care for stealth at all (but stealthy behaviour for AI is not working in this game so...).

There is so many things to fix with the AI that it should be in an expansion pack itself dealing with just it.

Edited November 11, 2013 by Bouben

[progamer 14]

This ticket here is related to the AI's navigation skills in a pitch black environment and flashlight use: http://feedback.arma3.com/view.php?id=12363

[-Coulum- 35]

Just thought I'd bring this too peoples attention for those interested in how AI function. Its kind of hidden in the editing forum and I personally find it very interesting and useful. Kudos for gammadust.

http://forums.bistudio.com/showthread.php?161681-AI-vs-AI-contest-(How-to-evaluate-AI-effectiveness-)

[oldy41 61]

For me the best way to fix the "terminator" AI, the random headshots and the way too short fights is to make the AI more arcade. What i mean is:

- Remove the recoil for AI (Since the main problem is always the first shot, who is the most accurate, recoil is pretty much useless)

- Increase the AI's gun's spread (In Arma 3, as it is in Arma 2, the weapons shoot pretty much straight and that's the main problem i think)

...

These rather simple tweaks might really help a lot while requiring moderate implementation effort.

In particular, increasing AI gun spread dynamically based on the situation would allow to implement things like fire for suppression rather easily without the need to implement more complicated (and cpu intensive) concepts like area targets.

(If an AI looses line of sight to an enemy just increase its gun spread temporarily and keep firing to the last known or suspected position of the opponent for a while. Voilá!)

I believe Enricksolt's observation really hits a nail: Some of ArmAs traditional problems are actually caused by the engine trying to be too exact in some aspects while not being able to keep up with others.

E.g. the AI is burdened by having to deal with weapon recoil, which it can not control in a reasonable way like players can. This makes AI machine gunners spray bullets with little chance to hit except with the first round.

Of course recoil effects should be simulated for the AI, but in a way which leads to realistic and immersive fire patterns. Abandoning AI recoil and instead tweaking gun spread dynamically might be a very creative, cheap and effective solution.

[xendance 3]

These rather simple tweaks might really help a lot while requiring moderate implementation effort.

In particular, increasing AI gun spread dynamically based on the situation would allow to implement things like fire for suppression rather easily without the need to implement more complicated (and cpu intensive) concepts like area targets.

(If an AI looses line of sight to an enemy just increase its gun spread temporarily and keep firing to the last known or suspected position of the opponent for a while. Voilá!)

I believe Enricksolt's observation really hits a nail: Some of ArmAs traditional problems are actually caused by the engine trying to be too exact in some aspects while not being able to keep up with others.

E.g. the AI is burdened by having to deal with weapon recoil, which it can not control in a reasonable way like players can. This makes AI machine gunners spray bullets with little chance to hit except with the first round.

Of course recoil effects should be simulated for the AI, but in a way which leads to realistic and immersive fire patterns. Abandoning AI recoil and instead tweaking gun spread dynamically might be a very creative, cheap and effective solution.

I'm pretty sure that currently the recoil for AI is simulated as increased spread like you said.

[_William 0]

In particular, increasing AI gun spread dynamically based on the situation would allow to implement things like fire for suppression rather easily without the need to implement more complicated (and cpu intensive) concepts like area targets.

(If an AI looses line of sight to an enemy just increase its gun spread temporarily and keep firing to the last known or suspected position of the opponent for a while. Voilá!)

Increasing spread isn't a good idea from a game design perspective, since (even if it were realistic) it increases the probability of the AI accidentally killing the player. Just witness this discussion thread to see how sensitive people are to being killed by the AI without being spotted by the AI, and without receiving warning shots (near misses) first.

There is plenty of know-how published on how to do plausible suppression fire (for example, GDC 2005: Killzone's AI ...(slides 26-34)), where being killed is more acceptable.

Keep in mind that (perceived) AI behavior is like food/wine: there is no accounting for taste.

William

[lordprimate 159]

that killzone stuff was good reading ... and that is a console game?... right? funny how a console game can handle those computations but Arma3's engine cannot... kinda says something about arma 3....

[progamer 14]

that killzone stuff was good reading ... and that is a console game?... right? funny how a console game can handle those computations but Arma3's engine cannot... kinda says something about arma 3....

Look carefully, the console game or other game will try to hide the problems it has. Nothing is perfect, atleast arma doesn't try to hide or disguise its problems.

[xendance 3]

Look carefully, the console game or other game will try to hide the problems it has. Nothing is perfect, atleast arma doesn't try to hide or disguise its problems.

Welcome to game dev, it's always about smoke and mirrors.

that killzone stuff was good reading ... and that is a console game?... right? funny how a console game can handle those computations but Arma3's engine cannot... kinda says something about arma 3....

High level AI isn't really computationally demanding. Except for those ray casts for line of sight checks Killzone uses.

[Make Love Not War 10]

High level AI isn't really computationally demanding. Except for those ray casts for line of sight checks Killzone uses.

Also don't forget pathfinding - not exactly computationally trivial either.

[xendance 3]

Also don't forget pathfinding - not exactly computationally trivial either.

Regular A* isn't that hard. It might take a while if you have lots of nodes though. But usually pathfinding tasks are deferred so that they're computed over a couple of frames, so that they don't stall the whole game.

[dmarkwick 261]

Regular A* isn't that hard. It might take a while if you have lots of nodes though. But usually pathfinding tasks are deferred so that they're computed over a couple of frames, so that they don't stall the whole game.

Perhaps indoors this might be true, but there are no defined nodes outside. The AI has to navigate a nodeless environment.

[xendance 3]

Perhaps indoors this might be true, but there are no defined nodes outside. The AI has to navigate a nodeless environment.

Yea, I was talking about Killzone. I just mean being a console game or not means nothing.

[Smurf 12]

Wild question: Rocket mentioned that zombie AI in DayZ aren't using nodes anymore but an "area scan" just to avoid obstacles. This couldn't be used indoors?

[oldy41 61]

Increasing spread isn't a good idea from a game design perspective, since (even if it were realistic) it increases the probability of the AI accidentally killing the player. Just witness this discussion thread to see how sensitive people are to being killed by the AI without being spotted by the AI, and without receiving warning shots (near misses) first.

There is plenty of know-how published on how to do plausible suppression fire (for example, GDC 2005: Killzone's AI ...(slides 26-34)), where being killed is more acceptable.

Keep in mind that (perceived) AI behavior is like food/wine: there is no accounting for taste.

William

I totally agree that the approach for using artificial bullet spread to implement supressing fire would not exhibit the _most_ clever of all possible AI suppression behaviors. Still I believe it is a very good trade-off between effect, implementation effort and cpu usage. (And please note that I did not suggest to target suppressing fire on the _actual_ position of suppressed units but either on their last known or the predicted current position. This would be totally valid, since it is close to human behavior.)

Also, I don't think the suppression algorithm presented for Killzone would be a good choice for ArmA simply because of the completely different map layout with no precalculated waypoint positions. Calculation effort for finding potential enemy firing positions would be much higher, since it would not be sufficient to check a limited set of nodes.

That being said I would be one really glad player enjoying every second of being suppressed by clever AI like that _if_ this can be achieved with reasonable cpu impact. :D

[maturin 12]

Killzone AI is completely irrelevant to ArmA. It can't be applied to a world larger than a shoebox with two exits.

[Make Love Not War 10]

Yea, I was talking about Killzone. I just mean being a console game or not means nothing.

Well, yes, for Killzone pathfinding probably wasn't that intense, but my comment was aimed at the huge, open world with multiple environment types (forest, field, urban, etc) that is ARMA. Also, don't forget that we need to account for not just infantry, but also a multitude of vehicles that operate on land, sea and air and at wildly varying speeds. Oh, and don't forget we also have to dodge around all these moving units and objects as well. Clearly calculations are going to be defered on-the-fly over several - if not many - frames. :) That said, it sure would be really awesome if somebody at BIS could sit down and rewrite the indoor pathfinding from the current handful-of-nodes-per-building state to something a little more useful.

Wild question: Rocket mentioned that zombie AI in DayZ aren't using nodes anymore but an "area scan" just to avoid obstacles. This couldn't be used indoors?

I could be wrong, but for indoors a simple polgonal surface mesh would be more than adequate. I was even tooling around with the idea of modding it earlier this year. From a conceptual point of view, the pathfinding itself wouldn't be too hard to write, but with the limited AI scripting tools available to modders, trying to integrate it with the underlying engine is a huge and hacky problem.

As for the scanning method mentioned for dayz zombie AI: one of the problems with a dynamic, real-time scan type pathfinding is that I expect it's difficult to plan any long distance routes with it. It's not an issue for brainless undead that just need to dodge around some obstacles so that they can swarm a nearby detected target, but how well would that work for a team leader trying to plan an ingress route into a nearby town? If anybody with some expertise could add or correct these thoughts it would be greatly appreciated.

Edited November 14, 2013 by Make Love Not War

[maturin 12]

Really, we just need buildings to have a node showing AI where the doors and windows are, with maybe a couple in the middle of rooms. If they would just move sort of mindlessly through shooting vantage points, it would be enough.

[gammadust 12]

Since we're in pathfinding, let me plug in the ticket:

0015297: Ability to dynamicaly set searchPath parameter values (via scripting or command menu)

[tmp95 16]

Really, we just need buildings to have a node showing AI where the doors and windows are, with maybe a couple in the middle of rooms. If they would just move sort of mindlessly through shooting vantage points, it would be enough.

+ 1 - Agreed. A somewhat simple yet effective for most part solution that would go a long way from where we are now. AI really is starting to hurt this game when one reads around the gamming forums (here and other places).

[ProfTournesol 956]

+ 1 - Agreed. A somewhat simple yet effective for most part solution that would go a long way from where we are now. AI really is starting to hurt this game when one reads around the gamming forums (here and other places).

Mkay, and what other games' AI can do the same as ArmA AI on such an open world ?

[froggyluv 2130]

Mkay, and what other games' AI can do the same as ArmA AI on such an open world ?

Hehe, although this is true and Ive probably posted that reminder more than anyone, I can't help but feel disappointed at the lack of evolution here. The interior stuff has already been said here ad infinitum, but what about some class specific behavior? Imagine if sniper had his own set tactics such as always trying to maintain his inherent distance advantage or seeking higher ground or a tower or such (such as GL4 mod). If MG'er set nests on the second floor with a man or two assigned to watch his back. If outgunned fire squad adopted a full circle defense -things that would actually warrant a tactics guide to how to deal with them (No offense Dyslexi) and gave the game a deeper tactical feel on how to deal with them. Make this default behavior that can be turned on/off via switch as to not break missions. As it is now -the official campaign takes zero risks and just place a few haphazard enemies on route doing the routine drop to prone, return fire, stand up sit down dance. No inside enemies, no driving up, disembark and fire from vehicle as cover etc... Yes, improvements have been made such as faster rotation with a distance check etc.. I was just expecting much more for this new flagship.

I'm not speaking directly at you Prof just venting :p

[Make Love Not War 10]

Since we're in pathfinding, let me plug in the ticket:

0015297: Ability to dynamicaly set searchPath parameter values (via scripting or command menu)

Yes, that would be extremely useful to have; already voted for it some weeks ago.

...AI really is starting to hurt this game when one reads around the gaming forums (here and other places).

Mkay, and what other games' AI can do the same as ArmA AI on such an open world ?

Let's be clear, from a technical standpoint, ARMA AI is hands down the best and most complex AI of any shooter out there that does - or, quite often, attempts to do - amazing things that AI in most other games doesn't even dare touch with a 30m pole. There is, however, a problem of perception by the end user. Namely, there is a major cognitive bias when it comes to AI whereby if it acts in a manner that the player deems to be human-like, the artificial actor in question almost completely escapes the user's attention. Provided everything and everybody within the virtual world continues to function in a way that approximates our expectations of how things should look and behave in reality, then the brain simply filters out all of this stuff and we literally forget about the technology that is working feverishly behind-the-scenes to make the illusion possible. But, as soon as we see something we perceive of as alien, artificial or out-of-place, our brains start freaking out. In regards to the AI, this means that the better and smarter the behaviour of the AI, the more it is ignored. It is only when the AI does something strange or stupid that the player notices, rages and then goes on an online posting spree. You have to understand that this noticing of only the negatives and flaws in the simulation is a natural human cognitive reaction, and is largely legitimate from the end user's perspective.

In this sense, one also has to appreciate that the BIS AI programmers are burdened with what is pretty much a Sisyphean task. Because, no matter how amazing the ARMA AI is going to get, there will always be problem areas, and so there will always be a continual torrent of complaints and little to no praise for what has been accomplished. How many people have noticed the major improvemnts in A3 AI over A2, for instance? In fact, there have been significant improvements over the last month or two, but because almost none of this is changelog'ed it sadly slips under the collective radar. All that is not to say that players and modders shouldn't be asking for more, but a little perspective helps sometimes.