AI Discussion (dev branch)

[Make Love Not War 10]

Oukej, thank you for confirming that the new presets won't take away any AI customizability from the end user. I was going to write quite a bit more, but there's already been too much recent ugliness in this thread. I'll just leave off by saying that the smartest way for BIS to fix the AI problem is to concentrate their efforts on giving scripters the tools to exert much finer control over individual AI units. ATM, many critical components of the AI are accessible to scripters either globablly, on a per group basis or are constantly overidden by engine internals. Expose a lot more of the AI framework within the SQF language and reap the benefits of crowdsourcing the problem.

[Variable 322]

^^ not as clear cut as you put it. While what you say is intuitive and true (due to progressive adjustment of aiming) you must not forget the competing factor which varies counter to that due to recoil.

Recoil should affect time between each shot - large recoil should make the AI taking more time between each shot, that's it. It should not affect the accuracy in this context. It is simply ridiculous to make the accuracy decrease as the AI keep on shooting. AI should have the exact opposite behavior. Sorry, but it's a clear cut case.

[fabrizio_t 58]

I'll just leave off by saying that the smartest way for BIS to fix the AI problem is to concentrate their efforts on giving scripters the tools to exert much finer control over individual AI units. ATM, many critical components of the AI are accessible to scripters either globablly, on a per group basis or are constantly overidden by engine internals. Expose a lot more of the AI framework within the SQF language and reap the benefits of crowdsourcing the problem.

This.

[enex 11]

Recoil should affect time between each shot - large recoil should make the AI taking more time between each shot, that's it. It should not affect the accuracy in this context. It is simply ridiculous to make the accuracy decrease as the AI keep on shooting. AI should have the exact opposite behavior. Sorry, but it's a clear cut case.

Exactly.You take a aim, shoot, you miss, you adjust. From worse to better.

[shark-attack 2]

^^^^ +1

[Fap 1]

Recoil should affect time between each shot - large recoil should make the AI taking more time between each shot, that's it. It should not affect the accuracy in this context. It is simply ridiculous to make the accuracy decrease as the AI keep on shooting. AI should have the exact opposite behavior. Sorry, but it's a clear cut case.

^ this x100

kju also mentioned the thing that bothers me the most, after AI reloads, the first shot afterwards is with aimbot precision.

Fix it. Just fix it.

Play your damn game and realize these things yourselves. These are so obvious.

I wanted to make a list of stuff that's broken, from the top of my head I could fill up 2 or 3 pages. With size 3 font.

[Gwynbleidd 28]

Recoil should affect time between each shot - large recoil should make the AI taking more time between each shot, that's it. It should not affect the accuracy in this context. It is simply ridiculous to make the accuracy decrease as the AI keep on shooting. AI should have the exact opposite behavior. Sorry, but it's a clear cut case.

Completely agree with this.

[gammadust 12]

Just an attempt at clearing my point:

I am saying that the more recoil the more negative impact in accuracy (red), and, the more time to adjust aim the more accuracy (green). These are competing factors. They compound in final resulting accuracy (blue).

It is true that the more time you give to compensate for recoil (as you compensate for aiming), it becomes less of an issue, but the more you bullets you fire the more time you need to mechanically compensate for each shot (the less time you have to aim also).

it is not as clearcut:

this is less of an issue if single shooting, as you should, more of an issue if bursting, even more if suppressing.

And this is still terribly simplified. ;)

[icebreakr 3157]

oukej: would be possible to set the accuracy and aiming flags in the Mission info and incorporate that into the mission file sqm itself? So the mission maker would decide how well "trained" is the waring side (well for each of 3 sides separately). That would be ideal, maybe not 3 settings but four:

- Recruits (20% accuracy, slow aiming speed)

- Regulars (30% accuracy, medium)

- Trained Army (40% accuracy, faster)

- Professionals (50% accuracy, high) 50% is believed to be BIS default setting right now, correct me if wrong?

AI simply cannot turn like a robot for 90 degrees almost instantly, aim in a millisecond to torso and fire off a deadly shot from 700m with a regular ironsight weapon. Cannot.

[oukej 2910]

oukej: would be possible to set the accuracy and aiming flags in the Mission info and incorporate that into the mission file sqm itself? So the mission maker would decide how well "trained" is the waring side (well for each of 3 sides separately). That would be ideal, maybe not 3 settings but four:

- Recruits (20% accuracy, slow aiming speed)

- Regulars (30% accuracy, medium)

- Trained Army (40% accuracy, faster)

- Professionals (50% accuracy, high) 50% is believed to be BIS default setting right now, correct me if wrong?

The sub skill functions are not linear and expressing the qualitative nature of it in percentage can be misleading. The descriptive words (recruits, professionals) are the way to go even if everyone imagines something different under such word.

Correct me if I got it wrong, but what you ask for is to have a "semi-global", mission specific (but not game-wide nor individual unit-specific) set of - let's say - sliders to adjust all the AI units under each side, am I right? It would be definitely be a new feature and I am afraid that it would clutter the settings. Not prio, but it's a good idea that could perhaps deserve it's own ticket? ;) Shouldn't be a big deal though to script such a solution, that would adjust the skill/sub skill values of all the units under a given side.

For more feedback about the AI Configuration & Settings, please use http://forums.bistudio.com/showthread.php?169122-AI-Configuration-feedback

[el_muerko 0]

Do you not think 50% accuracy is too high, which is compounded by the AI's likelihood of seeing you first, real armies would love to have soldiers hitting their target first time 50% of the time, it's not realistic.

[Sniperwolf572 758]

I'd like to add another dimension to this discussion. Right now it's mainly about the guts (skill coefficients, spotting, shooting, decision making).

The AI pathfinding/movement itself could use an improvement. Since OFP it feels very artificial. The AI "steers" as if it was a vehicle instead of a person. Small improvement came with Arma 2's micro-AI where they started to lean and strafe to adjust to cover, but in general the pathfinding never seems deliberate. They change their minds very often as to which "node" to go to and what the best way to get there is which results in a lot of this type of "steering".

Even the lone AI going from point to point loves to either use the stop-turn-go forward-stop-turn-go pattern or when they are not stopping, they wave about. I believe giving the AI a sense of deliberate movement would improve how AI feels immensely.

For example, certain improvements that would make the AI appear more human-like when moving would include things like, not turning their side/back (strafing instead) to the enemy when moving very short distances (adjusting position in cover, moving to a corner to lean, etc) or when the path to a further location requires a minor sideways adjustment, avoid physically turning but strafing to correct instead (imagine what you'd do to avoid a single tree when moving through a forest in a certain direction, compared to what the AI does now).

[Smurf 12]

One of these days I was bored so I started to do some tests with the AI. I couldn't believe in what I saw....

This is the "repro":

I would go around the corner, shoot the enemy group (often killing one) and hide behind the construction. I did it about 20 times and the most common result was this or something similar:

They would go prone in the tarmac, and slowly (setacctime 6!) go waaaaaaaaaaaay around, always "covering" each other, often looking at the wrong direction and getting cover behind... the air?.

A gif of another run, even worst result. They just gave up!

__________________

This one got close (closets on all tests) and I shot him dead.........which scared the rest of his group and ended fleeing.

Best results was this, which killed me. Happened once.

Another good result was a grenade launched at me wich also killed me. But happened only once too.

So I tried with a bigger group....

[Smurf 12]

Thought that with a bigger group they would perform better, more flanking and all.

It went kind of good when they spotted me on a flashing glimpse passing by the window:

But most cases went like the small group:

They never assaulted me on a organizated manner. Never used some kind of supressing fire. Never used a pinch manouver. Once used explosives. Use of cover = prone on tarmac.

I had the intention to test more scenarios but seeing them struggling with one so simple made me give up.

Just a mere observation that I wanted to share.

Ai settings, if that matters:

skillFriendly=0.80002338;
	skillEnemy=0.76002342;
	precisionFriendly=0.64004201;
	precisionEnemy=0.5680421;

[MavericK96 0]

Yeah, I've seen a lot of similar behavior doing small tests in editor (mostly just to try out new weapons and whatnot). A lot of the time when I engage a group and kill maybe one guy, the AI will just run back and forth, or retreat completely. Sometimes they will just go prone and sit there as I pick them off one by one, hardly returning fire.

I don't want to get insta-headshotted by AI, but they could use a lot more work to make them suppress, use flanking tactics, and overall just be more of a threat.

I mean, in ArmA 2 there are mods like ASR AI and TPW Supression that make the AI SO much more advanced than what we are seeing in ArmA 3.

[dnk 13]

Recoil should affect time between each shot - large recoil should make the AI taking more time between each shot, that's it.

Well, except if they're suppressing, specifically with SAWs/etc.

What bothers me the most is that BI made the AI accuracy DECREASE as the AI shoots a target instead of exactly the other way around.

This definitely.

2.

Take a look at mods that introduce AI supression.

Why stop there? Basically, ASR_AI3 should be a part of vanilla. A decent number of the TPW stuff would help add immersion also (falling, but also LOS). I'm guessing you guys can't just copy and paste their code, but it's clearly possible to make the AI look and feel a lot smarter in a number of situations, and it's clearly possible for the AI to use buildings.

The biggest issues (imo) are that the AI sucks under fire, the AI are superhuman in accuracy, and the AI can't use buildings. All 3 things clearly can be significantly improved upon. If it's an issue of worrying about breaking some missions or campaigns, make certain behaviors toggleable (like entering buildings and running far for cover) through scripts.

I'm just tired of the "supersniper-fish-in-a-barrel" AI. Their only challenge is their most frustrating "feature": superaim.

[nomisum 129]

The biggest issues (imo) are that the AI sucks under fire, the AI are superhuman in accuracy, and the AI can't use buildings. All 3 things clearly can be significantly improved upon. If it's an issue of worrying about breaking some missions or campaigns, make certain behaviors toggleable (like entering buildings and running far for cover) through scripts.

I'm just tired of the "supersniper-fish-in-a-barrel" AI. Their only challenge is their most frustrating "feature": superaim.

this. a friend who i want to play arma3 with always asks: is the ai still as freaking super accurate on long distances or is it like OFP where i actually managed to survive longer firefights. maybe he is caught in nostalgia, but he has a point.

Gesendet von meinem iPhone mit Tapatalk

[turin turambar 0]

this. a friend who i want to play arma3 with always asks: is the ai still as freaking super accurate on long distances or is it like OFP where i actually managed to survive longer firefights. maybe he is caught in nostalgia, but he has a point.

Gesendet von meinem iPhone mit Tapatalk

I think the big change from OFP to what we have right now it's the much higher detail in terrain, grass, vegetation, lightning, etc all makes harder to see the enemies, in a way that the AI isn't impeded in the same manner.

[instagoat 133]

I mean, in ArmA 2 there are mods like ASR AI and TPW Supression that make the AI SO much more advanced than what we are seeing in ArmA 3.

Do you have quantitative and qualitative evidence for that, or is this just another "gut feel" claim?

All the mods do is change the AI, but their behaviour is still very much hit and miss. By now I think the perception by the community that AI mods are improving the AI to such a large degree is down to the placebo effect.

For example the AI picking up At weapons never worked for me reliably, AI often picking up weapons from halfway across the map or picking them up directly in front of enemies. They still reliably hid on the wrong sides of walls, stood up in the line of fire, and the biggest change that was noticeable at all was the larger engagement distance and their decreased precision. (Which made them much easier to deal with than Arma 3's AI, who are much better at hitting things than some of the modded AI setups.)

ASR is a bunch of very good Ideas scraping right along the edge of the things that are possible, I grant that. However, I also think that what's under the hood is the only thing that can really make a difference. Making things even more complicated and convoluted by adding a script system on top of what already exists is really making things more difficult than they need to be.

Edit: RE OFP AI: you're having nostalgia glasses on. OFP weapons didn't have dispersion and the AI was super precise with them. The "Headshotted by a farmer with an AK across 400 meters in the fog" line didn't arise with Arma, it was a problem with OFP to begin with (back then, really, a feature.). Mods like BAS and community standards like JAM then came up with High Dispersion (HD) guns to make firefights last longer. Go back and try the game, and you'll see that you can't get through any firefight without casualties, due to the extreme precision of the AI. Arma 3's AI is not as deadly as the OFP AI was.

Edited November 27, 2013 by InstaGoat

[.kju 3244]

InstaGoat you seem to have very wrong expectations - both on what community modders can do and are up to do, and what BI is intending to do.

If you haven't realized it, there will be no major changes to AI.

Even their current effort is just limited to improving the difficulty settings as for AI skill and the documentation. That's it.

Edit: What comes after, or klamacz may or may not be working on is another story. Just saying you shouldnt open up yet another huge topic.

OFP AI is not the benchmark either - it is WGL/ACE/ASR/GL/SLX AI.

Both in terms of AI tweaking (aka configs) and AI features (improved behavior).

Edited November 27, 2013 by .kju [PvPscene]

[Bouben 3]

;2565440']

OFP AI is not the benchmark either - it is WGL/ACE/ASR/GL/SLX AI.

Both in terms of AI tweaking (aka configs) and AI features (improved behavior).

Sorry, but I have to agree with InstaGoat on this. While all of the mods gave something with a potential to the game, all of them also suffered the same problems the vanilla has. It was just tweaked so that it felt different and occasionally did something really neat but still did not have much needed reliability and balance that is missing in the series.

For example so applauded ZeusAI was, for me, an absolutely unusable dissaster that completely destroyed stealth factor and made every AI unit run out of ammo (not to mention other game balancing stuff that will show themselves only after careful testing). However guys who enjoyed firefights on long distance praised the mod in unlimited fasion, because they used it in specific situations and their view was biased by it.

The core has to be changed. Otherwise we will have these one-purpose-only mods that will make something better while making something else worse.

However, I cannot hide my excitement about Fabrizio's AI mod he is currently developing. This is the first AI mod that looks it does things more right than the others. We will see.

[instagoat 133]

I have to make myself clear here: I in no way intend to diminish or downtalk the achievements of the AI modders. What I mean is this:

Every AI modder can only do so much, has his own scope, and only a limited means of gathering dedicated feedback. Even when employing entire communities to test, there will still be biases, inherent limitations to the design approach chosen, as well as priorities that cater to the specific desires of that community. Because Arma is a realism oriented community, these changes often will be done to the detriment of gameplay (ie: ACE wounding system, as an example.). This makes sense for a modding platform, built by the community for the needs of specific groups inside the community.

However, AI mods with their added complexity cannot replace the core, and they cannot fix the flaws of the core. Basic changes and documentation are not enough, in my opinion, which is why I hope the major flaws of the AI can be adressed as far as that is possible. Not inside the code, I actually hope to see the Arma AI be developed to the point of being something that is limited by technology, rather than inherent code limitations. Cutting edge, sciency stuff, so to speak. Maybe it would make sense for BI to contact AI centers at relevant european Universities to set up a cooperation?

At any rate, I think the AI needs -major- work. Arma is about the only gaming platform right now that has such a freeform AI, and there is so much potential still in there that could be unleashed, when the resources can be invested into the work.

[dnk 13]

ASR_AI3 is much better at getting cover and not acting like idiots under fire (save they don't throw smoke). I don't care about picking up weapons, but after testing that, tpwcas, some other A3 mod, and vanilla, the current ASR is miles ahead of the lot. It's not perfect by a long shot, but it's a big step forward for the typical scenario in COOP of "shooting fish in a barrel". Now the fish run out of the barrel, do a decent job of returning fire, and in general act less like fish and more like sharks. I love it, and just wish at least more servers used it.

Why is this typical scenario important? Because it's how most players online know the AI through COOP, either Annex, Insurgency, Domi, or BECTI. That's the most public face of the AI outside of heavily scripted small missions.

[Bouben 3]

ASR_AI3 is much better at getting cover and not acting like idiots under fire (save they don't throw smoke). I don't care about picking up weapons, but after testing that, tpwcas, some other A3 mod, and vanilla, the current ASR is miles ahead of the lot. It's not perfect by a long shot, but it's a big step forward for the typical scenario in COOP of "shooting fish in a barrel". Now the fish run out of the barrel, do a decent job of returning fire, and in general act less like fish and more like sharks. I love it, and just wish at least more servers used it.

Why is this typical scenario important? Because it's how most players online know the AI through COOP, either Annex, Insurgency, Domi, or BECTI. That's the most public face of the AI outside of heavily scripted small missions.

So you test AI only in multiplayer COOP scenarios? Have you ever heavily played with an AI squad? Leading the squad etc? That is important too. The heat of a multiplayer game can drastically bias your AI behaviour observations.

Can you please post a link where I can download the ASR_AI3? I could not find it. Thank you.

---------- Post added at 17:00 ---------- Previous post was at 16:58 ----------

I have to make myself clear here: I in no way intend to diminish or downtalk the achievements of the AI modders. What I mean is this:

Every AI modder can only do so much, has his own scope, and only a limited means of gathering dedicated feedback. Even when employing entire communities to test, there will still be biases, inherent limitations to the design approach chosen, as well as priorities that cater to the specific desires of that community. Because Arma is a realism oriented community, these changes often will be done to the detriment of gameplay (ie: ACE wounding system, as an example.). This makes sense for a modding platform, built by the community for the needs of specific groups inside the community.

However, AI mods with their added complexity cannot replace the core, and they cannot fix the flaws of the core. Basic changes and documentation are not enough, in my opinion, which is why I hope the major flaws of the AI can be adressed as far as that is possible. Not inside the code, I actually hope to see the Arma AI be developed to the point of being something that is limited by technology, rather than inherent code limitations. Cutting edge, sciency stuff, so to speak. Maybe it would make sense for BI to contact AI centers at relevant european Universities to set up a cooperation?

At any rate, I think the AI needs -major- work. Arma is about the only gaming platform right now that has such a freeform AI, and there is so much potential still in there that could be unleashed, when the resources can be invested into the work.

I agree.

[Smurf 12]

Have you ever heavily played with an AI squad? Leading the squad etc? That is important too. The heat of a multiplayer game can drastically bias your AI behaviour observations

That is a whole other history. Commanding AIs is, and always has been a PITA. They don't couple very well with humans because the interface we have to interact with them isn't that great.