AI Discussion (dev branch)

[nkenny 1057]

Crew jumping out of a burning vehicle ought to be suppressed to hell and back. Possibly the best demonstration of a suppressed state is  found in RHS flashbang grenades.  The flashbang forces the AI into one of a number of animations.  Being animation locked for a short duration neatly and clearly demonstrated that the soldier is incapacitated/suppressed. 

 

In regards to active suppressive fire. I've been using a system similar, but less clever than shown by @fn_Quiksilver. The change is dramatic.  When the AI engaged in periods of (abstract) area fire it shapes the battlefield and emulated the behaviour of human players. Particularly when vehicles engage in more indiscriminate area fire (with their powerful and accurate weapons) the effect is good.   An 'attackTarget'  command would be very helpful. 

 

-k 

[damsous 329]

Understand that in Arma 3 AI crewman react to a modern HEAT, so when a very very hot spike enter your tank (its like a fireball) all the fuel and ammo are burning... trust me no one will stay inside

[froggyluv 2136]

2 hours ago, x3kj said:

In general, having a semi-random delay on AI reactions to observed state changes would be a very beneficial thing to do. Like crew noting their vehicle is damaged and ejecting -> it should take a short time between vehicle state reaching "damage, please bail out of me" and crew actually jumping out. It would make AI feel alot more life like, instead of super-machine with artificially increased bullet spread. Mentioned this already in one of the tank topics i think.

 

 Yep as well as tied into AI crew skill level. Hell just about everything "random" AI related should be tied into one of their sub-skills so the game could actually play out *cough* strategically?

[fn_Quiksilver 1636]

3 hours ago, froggyluv said:

 

 Yep as well as tied into AI crew skill level. Hell just about everything "random" AI related should be tied into one of their sub-skills so the game could actually play out *cough* strategically?

 

Every AI unit should also be given a randomly generated variable which has nothing to do with skills and stuff. just a random number to use for delays and stuff.

[Alwarren 2767]

8 hours ago, damsous said:

Understand that in Arma 3 AI crewman react to a modern HEAT, so when a very very hot spike enter your tank (its like a fireball) all the fuel and ammo are burning... trust me no one will stay inside

 

Not the point. The point is, AI reacts robotic to everything. They hear a shot, microseconds to turn in. Their tank gets hit, all of them jump out at the same instance. 

[damsous 329]

They react robotic cause they are (a computer manage the AI) you guys in arma want put 200 AI that react like human, and crying about performance....

Look at the other game Fallout for exemple.... there not more than 15 AI fighting.

In Arma you can get a good experience with AI if evrything is done with the editor, a mission with a spawn script will put a group with few waypoint, so a spawned mech group will do something stupid, a mech group managed via the editor if you doing right will disembark the soldier and cover the progression with suppresive fire....

 

"They hear a shot, microseconds to turn in"

you just need to set up the skill :

-AI skill : 60

-AI precision : not more than 50

 

for make it simple skill are spotting and engagement skill, so at 75 they will see you at 800 m if you are hide in a bush on the first shot, precision 50 they have 50% chance to hit you when they fire on you

so you must play with that setting + the AI skill in editor

 

[rübe 127]

8 hours ago, fn_Quiksilver said:

just a random number to use for delays and stuff.

 

A (small) random component is certainly needed, otherwise a crew with same settings still act all the same/like robots. Still, the main components should be: 1) latent skills (e.g. spotDistance), and 2) current conditions (e.g. fatigue or morale).

 

On a side note: does fatigue influence other skills? E.g. does spotting distance get worse with increasing fatigue? That'd be kind of cool, no? Anyways...

 

And then there's always the question of modelling various forms of communication/communication channels (e.g. radio). How long does it take for unit A to alert his leader (unit B) somewhere else, who in turn might alert officer (unit C). And can you kill unit A before that chain is set off?

And can we model the classic inversed pyramid of information in (mil.) communications? E.g. the most important thing first (we're being fucking attacked), giving more and more information/detail with more time/later (who/what is attacking? where is the attack? from what direction? casualties? etc.). The result being: the longer unit A can report about some incident, the more precise the knowledge for his peers will be (e.g. knowsAbout).

The problem here is that parts of this is hidden in the depths of the arma engine we can not influence, while the remainder can be easily scripted on top of it; but should we really have to, or wouldn't a native, maybe parametrized solution (maybe with triggers/some hook methods) be more appropriate?

[fn_Quiksilver 1636]

1 hour ago, rübe said:

 

A (small) random component is certainly needed, otherwise a crew with same settings still act all the same/like robots. Still, the main components should be: 1) latent skills (e.g. spotDistance), and 2) current conditions (e.g. fatigue or morale).

 

On a side note: does fatigue influence other skills? E.g. does spotting distance get worse with increasing fatigue? That'd be kind of cool, no? Anyways...

 

And then there's always the question of modelling various forms of communication/communication channels (e.g. radio). How long does it take for unit A to alert his leader (unit B) somewhere else, who in turn might alert officer (unit C). And can you kill unit A before that chain is set off?

And can we model the classic inversed pyramid of information in (mil.) communications? E.g. the most important thing first (we're being fucking attacked), giving more and more information/detail with more time/later (who/what is attacking? where is the attack? from what direction? casualties? etc.). The result being: the longer unit A can report about some incident, the more precise the knowledge for his peers will be (e.g. knowsAbout).

The problem here is that parts of this is hidden in the depths of the arma engine we can not influence, while the remainder can be easily scripted on top of it; but should we really have to, or wouldn't a native, maybe parametrized solution (maybe with triggers/some hook methods) be more appropriate?

 

remember arma is a game, “fun” is paramount. modelling these complicated things would use a lot of CPU on things a player may not know or care about, while the opportunity cost of that CPU time could have been used on something more “fun”.

[rübe 127]

@fn_Quiksilver: these are all punctual decision mechanisms, nothing that has to be run/checked every frame. And this stuff is really cheap to read out/compute/evaluate too, so I doubt a more elaborate model of these things would hurt overall performance in a meaningful way. Besides, I think that a more "human-like" behaviour (especially on the tactical/reaction side of things) can easily add a lot of value to the "fun" departement.

[froggyluv 2136]

2 hours ago, fn_Quiksilver said:

 

remember arma is a game, “fun” is paramount. modelling these complicated things would use a lot of CPU on things a player may not know or care about, while the opportunity cost of that CPU time could have been used on something more “fun”.

 

 For some of us, the lacking of these very pertinent military behaviour modifiers -ruins our fun.  I love Arma, but without heavy scripting interference AI battles play out way to haphazardly and samey when it just doesnt  need to be this way. My best example would be when I was playing a CTI a few years ago and saw the enemy had sent some Recon level troops toward the woods to the right of my base -for fun i sent a team of Repairmen and Pilots and tagged along to watch the massace (of my men). Results were anything but -they fought exactly the same and actually my guys (about 12 vs. 12 )actually got the better of them.

 

Now wouldnt it have been more interesting if my guys were actually frightened and it effected them? Knowing your going up against Force Recon in the woods while you should be changing a fuel pump?  The point is, incorporating a crew's skill level to time to bail out would be pretty low cost and a reasonable outcome on an albeit simple level. If i send my crack armor tank team in, they arent going to be pulling any Rommel level maneuvers angling for *cough* best angle of attack on enemy armor while expertly navigating the terrain -so why not at least make them more survivable then a Green level troop?

[damsous 329]

5 minutes ago, froggyluv said:

 

 For some of us, the lacking of these very pertinent military behaviour modifiers -ruins our fun.  I love Arma, but without heavy scripting interference AI battles play out way to haphazardly and samey when it just doesnt  need to be this way. My best example would be when I was playing a CTI a few years ago and saw the enemy had sent some Recon level troops toward the woods to the right of my base -for fun i sent a team of Repairmen and Pilots and tagged along to watch the massace (of my men). Results were anything but -they fought exactly the same and actually my guys (about 12 x 12 )actually got the better of them.

 

Now wouldnt it have been more interesting if my guys were actually frightened and it effected them? Knowing your going up against Force Recon in the woods while you should be changing a fuel pump?  The point is, incorporating a crew's skill level to time to bail out would be pretty low cost and a reasonable outcome on an albeit simple level. If i send my crack armor tank team in, they arent going to be pulling any Rommel level maneuvers angling for *cough* best angle of attack on enemy armor while expertly navigating the terrain -so why not at least make them more survivable then a Green level troop?

Force Recon are not super man, 2 regular soldier with terrain advantage and heavy weapon can easily annihilate a Force recon unit. A good exemple with the french Foreign Legion (2e REP) in Mali few men with Kalashnikov defeat an entire section (around 40 men) the Legion section lost few soldier and retreat....

 

Helicopter pilot and crewman are not always in a tank they also do basic infantry training...

[Alwarren 2767]

3 hours ago, damsous said:

They react robotic cause they are (a computer manage the AI)

 

No shit, Sherlock

 

3 hours ago, damsous said:

you guys in arma want put 200 AI that react like human, and crying about performance....

Look at the other game Fallout for exemple.... there not more than 15 AI fighting.

 

What does Fallout have to do with it? And who the hell was talking anything about performance?

 

But look, I get it. You are one of these White Knights that always pop up when someone says something in Arma could use improvement. It doesn't matter whether it makes sense or not, you just have to contradict. Then you have to resort to hyperbole because you have no serious point to make, and as a result derail the discussion. 

 

But that is completely beside the point. In Arma 2, Ai fired shots at the same interval. Sounded like robots. Then they added a variance in fire frequency, and immediately, the robot effect was gone. The same was true for AI pilot countermeasures that always reacted immediately when a missile was launched. It is neither unreasonable, nor in any way problematic, to have a small delay between AI perception and reaction. Nowhere does the number of AI come into play, nor if they behave like humans. They just should not behave like robots.

Tell me again, what's wrong with that?

[Varanon 892]

4 hours ago, damsous said:

They react robotic cause they are (a computer manage the AI) you guys in arma want put 200 AI that react like human, and crying about performance....

Look at the other game Fallout for exemple.... there not more than 15 AI fighting.

In Arma you can get a good experience with AI if evrything is done with the editor, a mission with a spawn script will put a group with few waypoint, so a spawned mech group will do something stupid, a mech group managed via the editor if you doing right will disembark the soldier and cover the progression with suppresive fire....

Uhm, you are missing the point. We're not talking about making the AI more sophisticated, just more life-like by adding a random delay to their actions.

Like, you remember some time ago, when AI was shooting, they go like "bang bang bang", like a clock. BI added a bit of a random delay before each shot, resulting in a more human sounding. You can apply the same principles to other areas of AI reaction. That's all we're talking about here.

So, I don't get why you are talking about performance (this has NO impact on performance whatsoever... well, yeah, in the nanosecond area). Also, even though beside the point, a mechanized group will behave the same, whether they are spawned with a disembark waypoint, or "managed" via the editor. 

So, I get the impression you don't really get what people were aiming at and just had to write something in defense of Arma just for the sake of it.

[damsous 329]

For exemple take few mods that will add few thing you want in game that result to a performance drop, for exemple GL5 or Vcom + AI first Aid script thats cool the AI suppress, get in the vehicle, use static, call mortar, they can heal and manage first aid on other AI or player, they will react and communicate each other.... all that thing without a single wp or trigger.... you can use it if you edit a really strict mission not more than 25 AI simulated... After on other gamemode like Liberation or other CTI or large scale coop thats become unplayable so a really smart AI will result to really poor performance gameplay so a bad experience.

 

And if you want talking about the improvement the little little improvement they do like pathfinding or other thing got a huge impact on performance, just compare 150 AI fighting in a town in Arma 2 and do the same in Arma 3. Its enjoyable in A2 but horrible in A3 cause too bad the performance.

 

About countermeasures AI should use it when they are locked by a AA.

 

 

[damsous 329]

After all i m not against your idea, but the game must be playable.

 

The only thing they can do with the actual state in arma 3 its to add new Waypoint, for exemple a waypoint for clear a house or something like this

[Tankbuster 1746]

@damsous You suggest that adding a tiny delay might have performance implications and now you're suddenly onto waypoints?

 

Whatever you're drinking - I want some.

[x3kj 1247]

Quote

you need to read before comment

ironic. We talked about delay for actions to mimick human reaction time. You go off on a tangent and accuse us of hypothetically whining about super AI not performing well because you somehow got the idea that waiting is a super complex task?...

 

3 hours ago, damsous said:

Im not talking for the delay only, you need to read before comment, pathfinding, delay and other thing all this change will affect all the AI....

which is precisely the idea. Give AI actions slightly random reaction time (aka delay) that are somewhat on human level, not supermachine level. There is no additional computational efford worth speaking of involved in that whatsoever. And no, fiddling with Skill sliders does NOT change reaction time for a vast amount of things (because there is none). Do you think we're morons and haven't tested stuff?

 

On 1.3.2018 at 12:45 AM, lex__1 said:

If the AI will spend more time getting out of the tank, he may have time to shoot at you. [...] The shooter still manages to kill you.

How dare he not jump out immediately as soon as track or engine is red...   .For the duration of delay time (between decision and execution), AI processes can be simply suspended -> because it's "in transition" out of the vehicle.

[twistking 204]

back to topic:

 

so, in order for turned out crew to feel less like a small-arms-magnet, maybe it would indeed be enough to implement a short semi-random delay until AI reacts to crew state change (turn out / turn in).

i think this is a very reasonable wish for an improvment, even so close to data lock...

 

this delay however would not scale very well with distance between AI and tank. so for units some hundreds meters away, a delay of few seconds would probably work, but if there were AI in direct vicinity to your tank (and f.e. comamnder would turn out to throw a grenade), the AI would be expected to react quicker to crew turning out.

so i still think that seperate spotting for vehicle and it's crew (while turned out) would be the most elegant solution.

 

another "quick and dirty" solution, would be to have the AI ignore crew up to a certain distance, unless AI is fired upon by said crew (FFV). so AI would still engage the vehicle with small arms (if possible) but would engage vehicles crew only within f.e. ~100m or when returning (FFV) fire.

 

any thoughts?

[x3kj 1247]

8 minutes ago, twistking said:

any thoughts?

what is your reaction time to a crew turning in/out? Does it increase if you go further away from the tank?  I would think not...

You are suggesting making things simply more complicated for one specific thing, as opposed to a simple solution (like reaction time delay across the board) for many relevant cases of AI reacting too quickly

[twistking 204]

12 minutes ago, x3kj said:

what is your reaction time to a crew turning in/out? Does it increase if you go further away from the tank?  I would think not...

You are making things simply more complicated for one specific thing, as opposed to a simple solution (like delay) for many relevant cases of AI reacting too quickly

 

well in this case, it would be less about "reaction time", but maybe more about a crude approxiamtion of spotting time. this is only about the "crew turn out problem", where crew that turns out (ffv) gets shot at immediately, because the vehicle is already spotted and the crew status change (turn out) gets reckonized immediately.

 

for most other cases, the semi-random delay would work very well and i would appreciate it being added.

my main concern that kinda started this recent discussion was about this specific issue though. but maybe the simple delay would work well enough as approximation. would need testing, i guess...

 

On 27.2.2018 at 4:38 AM, twistking said:

[...]

1. AI is super aware of crew being turned out and immediately concentrates small arms fire on turned out crew / holds fire imemdiately after crew turn ins.

-> AI should need a certain time to understand that crew is turned in/out. Ideally AI should be hesitant to engage tanks with small arms at all and maybe engage turned out crew only when in close proximity or when being engaged by small arms fire (ffv) themselfes.

[...]

 

[Alwarren 2767]

16 hours ago, damsous said:

Im not talking for the delay only, you need to read before comment

 

Wow, that's kind of ironic given the fact that delay was literally the only thing I was talking about before you brought up... 200 AI and Fallout.

[FallujahMedic -FM- 867]

Infractions have been issued and posts hidden. 

The majority may return to their civilised discussions. 

[x3kj 1247]

15 hours ago, twistking said:

well in this case, it would be less about "reaction time", but maybe more about a crude approxiamtion of spotting time.

Hm your objection is true. There is a difference. Reaction time for things i'm actively monitoring (display i'm watching shows big error message, radio i'm listening to starts transmitting, target i'm shooting does something) vs. perception time for things that change in my surrounding that are out of active focus. Reaction time is mostly consistent and short, perception time can be short or long - depends on "magnitude of change to my surrounding" and heading of that event to heading of my current focus and level of my current focus (e.g. tunnelvision)

 

[Strike_NOR 898]

Tunnelvision is one of the major things untrained combatants suffer from. It can be towards an active threat  (incoming fire, or advancing enemies), a point of interest (enemy position, vehicle, patrol squad) or while focusing/aiming at something.

 

One of the first things I was trained to do after firing at my target, is to scan left and right briefly before lowering my gun/self to see if there are more targets nearby. When reloading becomes automatic (you can do it without looking at your hands/weapon) you spend that time crouched (preferably behind cover) and look wide left and right to sustain situational awareness.

 

This way, you keep tabs on your squad, look out for threats and especially flanking maneuvers.

 

To draw this into ArmA AI, would in my opinion manifest in the form of narrowing the AI viewcone (scan) during engagements, and minimizing it further while firing, simulating the AI focusing. Anything outside the "focus" zone becomes increasingly more difficult to spot. As the time passes, the AI should gradually unfocus, and start getting a better peripheral view again.

 

The rate at which focusing and unfocusing happens, as well as the amount of time spent in "focused" mode should all tie into AI skill level.

 

Poorly skilled AI focus slowly, and remain focused long after losing line of sight to the target (ignoring everything else). When finally snapping out of focus mode, they unfocus slowly again.

 

This means that from the point of contact with enemy, they will struggle to accurately search for and find them, and once they do, they will not let sight of that enemy for a while. At this point they have tunnelvision and are unaware of other things happening outside the focused zone. They are basically "camping" the last known position of the enemy, instead of checking for other threats or relocating, making them vulnerable to flanking attacks.

 

It would be nice if AI had some delay to certain things. Such as reporting contacts.

 

A low-skill AI that happens to spot you at long distance, may halt and stare for a while, before reporting it to his group.

 

This means that you can tell when you have been made, and react before the entire enemy squad opens up on you.

 

There should be some delay to registering dead teammates in the same vehicle, like if turned out (or in) AI crew was killed. The rest of the crew would, depending on skill level, need a moment to react to the event and turn back in.

 

Perhaps this "delay" could be influenced by combat mode? "Safe" gives AI a delay increase (to simulate element of surprise), while "aware" and "combat" give decreased reaction time.

 

Same for registering when crew has turned in or out. Like others have mentioned, the AI seems to know instantly when crew turns in/out. This means turning out in the battlefield will get you killed instantly. Tried in vanilla, IFA and RHS. The AI, should need some time to register that you have exposed yourself, before taking shots at you.

 

Reloading could also be tied into this. Skilled AI recognize ammo count and reload when magazine has less than 15-20% ammo left and there is no line of sight with enemy. Unskilled AI don't care and will happily meet you with 1 round left in theie magazine (seen this so many times. AI see you, fire one shot and instantly starts reloading infront of you).

 

Just thinking here that this makes it possible to better simulate untrained combatants vs trained combatants in different scenarios.

 

 

Would be nice if the AI Tank gunners view cone (focus) was generally really narrow, the AI drivers was forward in a ~180degree arc (unfocused) and the AI commander was at a semi-focused ~45degree arc and regularly scanning 360 degrees to look for threats. This way, the commander would do the spotting, gunner would focus on enemy and shoot, while driver keeps situational awareness forwards.

 

Pair that with reaction time and you actually stand a decent chance against enemy AI in the field, and ambushes become more valuable in terms of gameplay.

 

 

[damsous 329]

Maybe there is something that can be fixed with the AI vs Armored, if the crewman behavior is "SAFE" (so their head are outside the tank) the AT specialist will use their rifle for shoot the crewman..... they will use the AT when the crewman are in danger or aware, after that they can't fire a missile cause they are spotted and killed by the tank.

So its easy for an AI to take down a tank in combat mode and impossible when they are safe cause they use rifle.

 

But really guys a forum its for debating a subject, debating=contradiction sometimes, but its like if somebody not agree something he's the public enemy.

 

I think arma need to fix few thing before adding other thing that can cause new issue.

 

Im not a dev or something like this, but when i see adding delay.... imagine 1 sec an AI jet pilot if you put more delay on reaction.... he will just fly around the target without engaging, same with a AA he will never able to shot down a fastmover.

Edited March 2, 2018 by damsous
faute d orthographe