The AI's reaction time/ability is insane.

[pd3 25]

I'm not trying to beat a dead horse here, but there are some thoughts I have to offer on this subject, I know I may get some salty remarks, but I would ask anyone disagreeing with me to remain civilized.

Something I've noticed in the more recent update is the propensity for the AI to immediately fire the second even a sliver of the player is visible, even at considerable distances. Even very narrow gaps between buildings if the player is moving quickly between them and the AI hasn't previously detected them, they will fire, and often connect. I honestly don't think I could identify, and respond to a target that quickly in those circumstances myself.

It forces the player to either assume a level of white knuckle robotic accuracy, which honestly I have no desire to do, I play because I enjoy the game, my Quake-era deathmatch days are long behind me. That sort of MLG tier effort you'd have to put into it is not what I'd consider enjoyable, it feels like work. No actual human being is going to move that quickly at any rate, it breaks the atmosphere of the game to an extent. People seem to overestimate human kinesthetic ability when they're playing games anyhow, call it a form of gaming dunning-kruger.

The AI reacts within minute fractions of a second, in situations that would spell death for most players attempting the same maneuver, the AI can simply wander around blind corners, face to face with the enemy and stand a very good chance of prevailing. All this forcing to player to employ tactics that assume the AI will kill you in a straight out fight, or in situations that would be deemed risky at best for the player.

Having to think and respond asymmetrically is fine and well, but it's made CQB not terribly fun. Unless you're willing to crank down their weapon dispersion or increase how shaky their aim is, which precludes them from making any aimed long range shots. You're pretty much having to assume the AI will be able to out shoot you in stand up fight quite a lot of the time. Similarly lowering the spot time so that it has an impact on CQB type situations makes the AI incompetent when spotting the player at any considerable distance. I don't really think any of those adjustments are reasonable as it breaks one part of the game in favor of improving another.

I suppose my main problem with the AI in this respect could be broken down into a few criteria.

- Their threat identification speed is so fast that I've now often seen the AI reactively fire at a wall upon first perception of a "target" (implying they aren't even aiming at the player) and then rapidly turning with CIWS-like precision and speed toward the target. This as opposed to most players who would have to at least see enough of the enemy to positively ID them as such or at least have a decent profile to shoot at. As mentioned, it seems as if they "see" a mere sliver of their target, they go mad and fire at the wall reflexively and quickly and accurately swing their aim toward the player. Expanding the AI criteria of what is necessary to constitute a "good shot" seems like a good idea to me.

- As mentioned, they will start shooting before they're even aiming at you, I've experienced this quite a few times now, it is a thing they do. I don't know how many players actually do this (I sure don't), and it seems ridiculous that the AI would do it. The AI should be subject to some conditions requiring it to actually have a decent "sight picture" on a reasonably sized portion of the target before firing (not the tip of the player's boot). Too many times they behave like psychic maniacs whose powers of perception work far faster than their bodies do, and that's saying something considering their "physical" reaction times are nothing to sneeze at.

- The AI seem to be able to "see" things they shouldn't. There have been a few times where the AI's "head" (presumably their origin of perception) was blocked by an obstacle, that is to say they were behind the corner of a wall, with only a bit of their leg barely sticking out. Regardless of this, they would fire the second I moved to glance at them, yes they would fire into the WALL, they could perceive my moving though technically having their vision obscured. From such an angle I might add, that it would be impossible for the AI to actually "see" me had they been a human player, although I could see a bit of their leg poking around the corner. This sort of "extrasensory perception" in the AI is becoming something of a drag and is taking a lot of the fun out of MOUT and CQB type engagements.

I suppose this might actually put a strain on the performance of the game, but making the amount of a target that the AI sees act as a coefficient of how quickly the AI identifies or reacts to a threat would probably help a lot. Possibly even making how much of the AI's target they see being a limiting factor to actually firing on the target might also be a good idea as well.

Perhaps these phenomena I'm describing aren't all that new, but something about the AI's perception and how they respond to that perception seems to have changed.


 

[en3x 209]

Interesting new things are coming up.

 

Credit to Cosmic for records.

 

 

 

 

 

[scottb613 285]

Hi Folks,

 

While brand new to Arma - I'm not brand new to FPS... I've been working my way through the first vanilla campaign and the CIWS analogy seems appropriate - LOL - and funny... I'm a huge fan of "realistic" and the AI seems to be super human unbeatable... I've noticed some mods in the lib to adjust the AI - do any of these work well ??? I've also tried adjusting the skill level in the settings - didn't seem to really change anything... 

 

The videos look MUCH better - and - nothing like my experience so far...

 

Regards,
Scott

[froggyluv 2136]

 My experience testing (at least with 1,50 RC) is the same as Cosmics. You can very much have stealth/ambush firefights given that the enemy AI is still in Safe mode -much harder when they are in Alert or beyond but hey now thats pretty much how it should be. Upon similar testing behind bushes with AI striding towards me maybe 10m from the side they would sometimes stop and freeze as if taking time to decipher something not being quite right -and then open up on me.

 

 This little window of time, is when much could be added to really flesh out the human response. Small animations like a cocking of the head to catch a sound or the like, or perhaps a squad instead of a lone soldier, the detecting soldier calls out or throws a 'Halt' hand signal up, or starts to point in your general direction with all heads turning to look -this would lead to some serious tantalizing gameplay experiences.

[pd3 25]

@en3x

I am really interested in that part in the first video at 1:50.

How much of the player do you think was visible, it doesn't seem as if it was much, and that lightning fast identification time is what I'm well accustomed to. If he had a line of sight to shoot you'd be in a spot of trouble for sure.

I personally find the speed at which the AI identifies obscured threats and the speed at which they can draw on them to be a little annoying, I don't care if other people find it to be fine, but I'd definitely like to have the option to scale it back. The video highlights some of the concerns I was describing in my initial post. You have to be super-asymmetrical or extremely fast on the draw to make an opportunity of such a situation.

It wasn't as bad as the AI seeing you on their periphery and wheeling about to put a bean in your helmet, but still potentially annoying.

I personally find this video the most telling about what I find irksome about the AI personally, however it's nice to know that smoke actually works.

Same with this video, in fact the AI does that thing where it shoots at the player and hasn't even properly aimed at him yet it seems, as you can see the shots correct. All way too fast for my liking, the AI are like super fast inhuman replicants. 

The time between the point of identification to the first shot going off is a bit on the low side for my liking, there is no implication of human error or delay. It reminds me of the old days of the original Rainbow 6, if anyone recalls that game as well.

 

 

"much harder when they are in Alert or beyond but hey now thats pretty much how it should be"

 

I don't necessarily agree with that, especially since I've seen the AI do insane shit even though I've moved out of it's visual range, It's one thing if the AI's position doesn't change and neither does the players, however once the AI loses sight of the player, that should leave the AI more open to being surprised again if they're not expecting the player to be in a given position while moving around. They're just too sensitive once they've been activated into something outside of safe mode. If an AI advances on my position and doesn't or shouldn't explicitly know where I am, I should by all rights have the advantage if they blunder around a corner unaware that I was there to begin with.

That just doesn't seem to be the case in MOUT/CQB type scenarios from what I've experienced.

 

[Grumpy Old Man 3546]

Same goes for AI controlled vehicles.

When put into safe mode they turn off lights as soon as the first bullet hits the vehicle with no delay or any sign of reaction time.

[dnk 13]

Beaten to death indeed, yet for some reason we keep having to play with these same AI...

 

It's an MP game now. It's really hard to claim this is a worthwhile SP game or even COOP. There's good communities for sure, but the opponents are ridiculous and unreal. As someone who likes to play against BALANCED AI, it's quite depressing. It's just not really an option in this game. All the mission makers seem unable to choose realistic settings for the AI, so we're left with laser-accurate instant-reaction killbots.

[fn_Quiksilver 1636]

@en3x

I am really interested in that part in the first video at 1:50.

How much of the player do you think was visible, it doesn't seem as if it was much, and that lightning fast identification time is what I'm well accustomed to. If he had a line of sight to shoot you'd be in a spot of trouble for sure.

I personally find the speed at which the AI identifies obscured threats and the speed at which they can draw on them to be a little annoying, I don't care if other people find it to be fine, but I'd definitely like to have the option to scale it back. The video highlights some of the concerns I was describing in my initial post. You have to be super-asymmetrical or extremely fast on the draw to make an opportunity of such a situation.

It wasn't as bad as the AI seeing you on their periphery and wheeling about to put a bean in your helmet, but still potentially annoying.

I personally find this video the most telling about what I find irksome about the AI personally, however it's nice to know that smoke actually works.

Same with this video, in fact the AI does that thing where it shoots at the player and hasn't even properly aimed at him yet it seems, as you can see the shots correct. All way too fast for my liking, the AI are like super fast inhuman replicants. 

The time between the point of identification to the first shot going off is a bit on the low side for my liking, there is no implication of human error or delay. It reminds me of the old days of the original Rainbow 6, if anyone recalls that game as well.

 

 

 

I don't necessarily agree with that, especially since I've seen the AI do insane shit even though I've moved out of it's visual range, It's one thing if the AI's position doesn't change and neither does the players, however once the AI loses sight of the player, that should leave the AI more open to being surprised again if they're not expecting the player to be in a given position while moving around. They're just too sensitive once they've been activated into something outside of safe mode. If an AI advances on my position and doesn't or shouldn't explicitly know where I am, I should by all rights have the advantage if they blunder around a corner unaware that I was there to begin with.

That just doesn't seem to be the case in MOUT/CQB type scenarios from what I've experienced.

 

 

Remember this is trying to simulate AI as it will be in 2035 :)

[Greenfist 1863]

 

I don't necessarily agree with that, especially since I've seen the AI do insane shit even though I've moved out of it's visual range, It's one thing if the AI's position doesn't change and neither does the players, however once the AI loses sight of the player, that should leave the AI more open to being surprised again if they're not expecting the player to be in a given position while moving around. They're just too sensitive once they've been activated into something outside of safe mode. If an AI advances on my position and doesn't or shouldn't explicitly know where I am, I should by all rights have the advantage if they blunder around a corner unaware that I was there to begin with.

 

Yes! This has been one of my biggest gripes with AI. If he loses LOS, I flank and appear next in his field of view (which is about 150 degrees I think), he turns and shoots pretty accurately in about 0.1 seconds. Just if I had re-emerged right in his crosshairs.

 

The whole "estimated target position" feature is cool, but it's completely meaningless in this kind of situation.

[fareast 20]

i think ai should at least randomly shoot the bushes if he has plenty of ammo..or throw a grenade maybe

[oukej 2911]

...or throw a grenade maybe

And so it does. If it's only estimating the target's position it's also quite prone to verify it by throwing some nades on it ;)

Ad reaction time - we've seem complains about it.

Even when the reaction is too fast it still should take the AI some time to aim on the target properly (and be able to score a hit). The AI aiming gets disrupted by any activity like reloading or turning around. Also the closer the target is the bigger tolerance there is for aiming off the target. If something pops up close to you you spray it and hope for some hits, if something is far away you may want to stabilize your aim first. AI does exactly the same.

[rübe 127]

Ad reaction time - we've seem complains about it.

Reaction time alone is not good enough, I'm afraid, or we'll end up with units looking somewhat ridiculous.

IMHO it's important to also have a visual cue about what's going on in the mind of an AI. Being alerted, alerting others in the same group (to look somewhere), maybe use the radio (if in safe and not already in combat mode), small gestures, looking around/observing, maybe taking a more defensive stance/position for a moment. Such little things.

[soulis6 24]

I think you hit the nail on the head with your first point, that enemy ID is just automagical and usually instant. Although to be fair, most games that aren't stealth focused(metal gear, splinter cell, etc) have this problem. This is the biggest glaring problem with the AI in my opinion, and if it were up to me, the first thing i'd try and make changes to.

 

Basically the problem is that there's no 'threat condition' or 'perceived danger' value with units. A routine patrol in a friendly controlled base next to a civilian city, where they may not have seen enemy contact for months will react just as fast to a half-glimpse of an enemy or a partial ID and will be just as quick to fire, as a fire team in hostile territory who were recently engaged and are expecting to see enemies around the next corner.

 

In both cases, just seeing unknown movement from the corner of their vision for a second will cause them to immediately open fire without warning their surrounding group or double checking that the threat is indeed an enemy, but really in the case of the routine patrol in a friendly base, you'd want them to maybe call out or move to get a closer look or report something suspicious to their squad leader.

 

I think a good solution would be to give AI groups an 'expected danger' setting, able to be set by the mission editor, and possibly changed during gameplay (if they move from a friendly base to the frontlines for example), along with a unit specific internal suspicion gauge/meter. This suspicion meter would go up faster or slower depending on this expected danger setting, whenever they see something out of the ordinary (i.e. sighting an unknown partially obscured unit, or seeing movement on the edge of their peripheral vision, etc), and when this meter reaches a threshold it would kick off some sort of behavior. If it was just a quick glimpse where they aren't sure they actually saw anything (the meter didn't reach this threshold), they would just resume normal behavior and the meter would slowly decrease. 

 

What behavior it kicks off is a whole other question, and not an easy one, given the amount of variables and different situations that could arise from this, but something easy at least would be to let the squad leader know they saw something unidentified over in X direction, so they could at least direct someone to watch that area. More advanced would be having someone go over and get a closer look if it's nearby, or yell a warning or identification request, etc. This behavior could be specified by the groups 'expected danger', as it would certainly change depending on situation. 

 

And all of this would of course be inactive while the group is in a active combat state, they just continue to act as they do now, being quick to identify and target hostile units.

 

But all of this would allow a lot more nuanced stealth and special ops gameplay, which to my mind is pretty essential considering the game is often geared towards "Behind enemy lines" type special ops missions, or nighttime raids, that sort of thing. It definitely wouldn't need to be as in depth as a full on stealth game, but just a few changes in the way that target ID works would go a long way I think.

[pd3 25]

All things considered, I want to point out that I really enjoy how the AI behaves in other ways.

I've had some very interesting experiences with enemy AI squads.

I had one circumstance in which I was fired upon in an urban environment at a distance, I took cover behind a wall, moved to the opposing side of the wall to get a look at who/how many were shooting at me from a different vantage point.

When it became obvious that I wasn't going to engage the AI at range, they actually left the building they were occupying and advanced on my last known location, the manner in which they did this was what impressed me.

Two advanced the most obvious route, however a third one went around the other side, whilst a fourth hung back. The three advancing attempted what was basically a pincer maneuver on me, which I thought was really clever. I LIKE that aspect of the AI.

What I don't like is their superhuman recognition/reaction time/ability, it just puts a damper on the experience.

 

 

In both cases, just seeing unknown movement from the corner of their vision for a second will cause them to immediately open fire without warning their surrounding group or double checking that the threat is indeed an enemy, but really in the case of the routine patrol in a friendly base, you'd want them to maybe call out or move to get a closer look or report something suspicious to their squad leader.

Yes, absolutely. Their peripheral vision is as good as a fully-exposed view of their potential target.
 

 

And so it does. If it's only estimating the target's position it's also quite prone to verify it by throwing some nades on it   ;)

Yes, I've noticed this as well, I actually have no problem with this, it's added a bit of a dynamic twist to close/medium range engagements.

Remember this is trying to simulate AI as it will be in 2035  :) 

[tortuosit 486]

Basically the problem is that there's no 'threat condition' or 'perceived danger' value with units. A routine patrol in a friendly controlled base next to a civilian city, where they may not have seen enemy contact for months will react just as fast to a half-glimpse of an enemy or a partial ID and will be just as quick to fire, as a fire team in hostile territory who were recently engaged and are expecting to see enemies around the next corner.

 

[...]

 

I think a good solution would be to give AI groups an 'expected danger' setting, able to be set by the mission editor, and possibly changed during gameplay

 

You describe the problem very well. And an interesting idea, worth some thoughts I think. Fractional values [0..1]... This variable, mixed up with number, and summed up time of clear sights on enemy... and distance to enemy... = delay until AI attacks. Some kind of attack stack. If it's full -> attack.

But tbh I guess something like this already exists :D

[oukej 2911]

Basically the problem is that there's no 'threat condition' or 'perceived danger' value with units. A routine patrol in a friendly controlled base next to a civilian city, where they may not have seen enemy contact for months will react just as fast to a half-glimpse of an enemy or a partial ID and will be just as quick to fire, as a fire team in hostile territory who were recently engaged and are expecting to see enemies around the next corner.

This comes down to the game as a platform and it's generic-ness. One premise can be that there's "always a war going on". I'm definitely not saying that it always needs to be like that, but making a different setting may become more difficult.

I'd recommend trying different AI sub-skills values.

You can also create a repro mission with an obviously unnatural reaction time of the AI so we can narrow it down.

[rübe 127]

It's indeed hard to improve (default) behaviour without compromising and losing generality. This wouldn't even be a problem if it weren't so hard to model different/extra behaviour on top of the default one.
 

I'd recommend trying different AI sub-skills values.

I'm sure that one can model behaviour with skills to some degree, but I doubt this will really cut it. Have you guys ever thought about introducing a set of AI rules of engagement (ROE) values, complementing the skills stuff?

 

Should units shoot unidentified units on sight, no questions asked? Or should they be more calm, only opening fire if needed/shot upon? How about active/aggressive vs. reactive/defensive combat behaviour on a scale from 0 to 1?

Should units engage enemies once identified? As in flanking and stuff? Or should they rather sit tight?

Does a group use the bounding overwatch dance or rather run and gun?

...

 

This would also be an opportunity to implement different (default) faction behaviour. Strict ROE and tactics for the well trained guys, not so much for guerilla dudes.

[pd3 25]

Unless the aimingSpeed values go into the hundredths, then even at a tenth it's still on the fast side.

[soulis6 24]

This comes down to the game as a platform and it's generic-ness. One premise can be that there's "always a war going on". I'm definitely not saying that it always needs to be like that, but making a different setting may become more difficult.

I'd recommend trying different AI sub-skills values.

You can also create a repro mission with an obviously unnatural reaction time of the AI so we can narrow it down.

 

That's true, and I definitely understand that there's so many different situations that make it really hard to make one universal solution, but like Ruebe says, it's more about the rules of engagement. I agree with all of what he said, it would be awesome to have some more control over the enemies training/instruction and their behavior while not in combat.

 

Tweaking the skills wouldn't really fix the fact that most of the time when you're sneaking around somewhere that's not on the frontlines they just don't really react with any restraint or sense. It's not an easy problem for sure, because as soon as you get into more nuanced reactions than just 'OPEN FIRE', it throws all kinds of potential problems in there; if the unidentified unit is too far away, or there's no clear path to investigate, or the group is busy with another command, and so on and so on. 

 

But even just a simple reaction where they see something nearby and stopping and looking over to where they see the unidentified unit or possible contact or whatnot, and waiting for a few seconds while they focus in that direction with an animation that is visibly showing them looking at where the contact was, would go a long long way.

 

I think no matter what, there's basically no reason for a unit that's in SAFE combat mode to immediately open fire without even confirming with their SL, or giving pause to confirm that what they're seeing is an enemy, unless of course their under fire or someone has just died nearby or whatever else normally sets them into danger mode.

 

The only sub-skills listed there that would tie into this, would be Spotting time and spotting accuracy. In my example of the 'perceived danger' stat, that stat would probably affect both of these sub-skills heavily, but it would just be one component i think that would be needed.

[fn_Quiksilver 1636]

group setCombatMode "Green";

 

and the units in the won't engage right away unless you're quite close.

 

contrast with a combo of:

 

group setBehaviour "SAFE";

group setCombatMode "RED";

 

and the units in the group will engage as soon as they see a target.

[pd3 25]

Interesting, I'm partial to using ALIVE myself, so I'm not sure how I could work that into use with it's dynamic spawning, but it's a work-around for mission designers.

I wonder if ASR_AI or any other AI mods alter that in some way so as to be useful to that end.

[fn_Quiksilver 1636]

A lot of AI modders set the combat mode to "RED" to make the AI more challenging. I do that too in most cases. The players usually come in dry, if you get my drift, so I buff the AI except for their aiming accuracy, which I generally leave quite low by default but allow it to adapt to certain variables.

[bad benson 1733]

Yes! This has been one of my biggest gripes with AI. If he loses LOS, I flank and appear next in his field of view (which is about 150 degrees I think), he turns and shoots pretty accurately in about 0.1 seconds. Just if I had re-emerged right in his crosshairs.

 

The whole "estimated target position" feature is cool, but it's completely meaningless in this kind of situation.

 

 to me this is the main problem. i can adapt to AI generally being very sharp sensed and stuff (or tweak with skill array stuff mentioned by oukej). but these kind of things are what kills it because there is no explaination like "they're well trained...there's a war going on" that works there to keep immersion intact. it simply makes no sense in any case that they still know.

 

i am always wondering what data is actually being used to create behavior and simulated knowledge (since the game just knows everything at all times you should be able to pick and choose) sometimes, especially since that knowsabout discussion in the AI thread a month or so ago and if it will ever change. because as you said. some of the data seems awesome but it doesn't reflect in the behaviors.

 

i'd go as far as saying that making their knowledge way faster in how it can change (concealed flanking example is perfect) would be a giant leap for arma's AI. i've seen some changes being made that go in that direction but it always feels just like multipliers put ontop of something that is wrong at the core.

 

they need a "search but don't know" state for when they lost direct line of sight until they get new input like a shot sound or new visual. i don't want idiots with alzheimer's but the whole thing where once you're spotted they enter that mode where they just seem to know and only multipliers (or whatever you wanna call them) are keeping them from coming straight for you keeps them from being very interesting to deal with.

 

i'm just interpreting the end result of course maybe it's already possible with tweaked values. i also wonder how much of the knowledge stuff is engine side and how much of it is in the FSM. because if the FSM is just taking certain things for granted and forcing them to appear different then maybe we need to go deeper or something.

[twisted 128]

It's indeed hard to improve (default) behaviour without compromising and losing generality. This wouldn't even be a problem if it weren't so hard to model different/extra behaviour on top of the default one.

 

I'm sure that one can model behaviour with skills to some degree, but I doubt this will really cut it. Have you guys ever thought about introducing a set of AI rules of engagement (ROE) values, complementing the skills stuff?

 

Should units shoot unidentified units on sight, no questions asked? Or should they be more calm, only opening fire if needed/shot upon? How about active/aggressive vs. reactive/defensive combat behaviour on a scale from 0 to 1?

Should units engage enemies once identified? As in flanking and stuff? Or should they rather sit tight?

Does a group use the bounding overwatch dance or rather run and gun?

...

 

This would also be an opportunity to implement different (default) faction behaviour. Strict ROE and tactics for the well trained guys, not so much for guerilla dudes.

 

This so much. The setskills are useful for a generic base set. But a some behaviours like mentioned here would allow them to perform more realistic and also give dramatic clues that the player could pic up on. Things that communicate a great deal to player that setskill cant.

[YanYatCheng 96]

AI was always broken, even COD4 did that better than ArmA3.

 

sniper are too damn accurate at close range (MLG 360 faze trickshoter that rekt you with one shot up close, never miss)

AI can see trough glass/trees

AI can headshot you no problem (nearly everytime)

AI will know where you are even it can't hear/saw you coming

 

just... OMFG, so unrealistic 

 

I am quite good in PvAI, I can get a long streak against AI, but really, do they even try to make AI realistic?? even a normal FPS game got better AI than ArmA3