ARMA 2 Micro AI Thread

[Estar 10]

How long were the firefights people had so far? For me the fights are still rather short and in concept very similar to the way we fought in OFP and ArmA.

The new AI shows some impressive abilities compared to ArmA but it seems mostly about moving close to cover, not to actually fight from cover. Kneeling next to a corner or tree is nice for looks but it provides as much cover as kneeling on open ground. As such it is my impression after playing Arma 2 for two weeks pretty intensively that combat is not fundamentally different than in ArmA.

There is AI that leans around corners or shoots over small walls but it happens not that often. Of course it has also to do with the fact that, save for urban terrain, there is not that much cover available in the world (as trees are no cover). The AI seems generally to be very reluctant to fight from good positions should they find one, moving further quite quickly. Also I do not have the impression that suppressive fire works for the AI on the receiving end. I do not say it isn’t in the engine but while playing I could not yet see it making any noticeable difference.

The concept of fighting from sheltered positions, exposing only as little as possible when firing and to quickly duck back to full cover when under fire, still seems to be missing from the AI. I hoped for longer firefights from clear positions in Arma 2 but my impression is that combat still plays pretty much the same as previously.

Yup, this is the exact same impression I have of the ARMA2 AI.

[Ketamine 10]

It always needs huge improvement (especially AI), but do you know any game where battles and troop (ai and player) numbers of this scale can run as well as ARMA2 in such a realistic environment?

Maybe if this small dev company had huge budgets, then the AI could be improved.

But do you really want the money people such as EA touching this beautiful game? I think not! You know what would happen. Look at OFP2.

What we can hope for is for the devs to take all the (decent) feedback and work on new uber AI code etc, to make it the true vision that it is so close to.

How long were the firefights people had so far? For me the fights are still rather short and in concept very similar to the way we fought in OFP and ArmA.

Not sure what game you where playing, perhaps your confusing ARMA with CS, but the longest ARMA1 game i played was about 14 hours. I only left due to the serious need of real life food and sleep. Clearly you have not experienced the true ARMA full map, battle experience....

Edited June 13, 2009 by Ketamine

[CarlGustaffa 4]

52 hours here, in a single game. That does however, include a nice little collapse over the keyboard :) In OFP the 'record' was 5 days, but then we played a massive coop campaign spanning over quite a few missions. However, these were game sessions, not firefirhgt. Longest firefight in Arma1 was about one hour against AI in a highly modified version of Domination, and we gained nothing but a hill crest (?). Once that crest was ours, we had little resistance doing the rest since the enemy forces was so thinned out. This was all using an early version of ACE, where we or the enemy couldn't really hit shit. If this was for real, I think the firefight went on pretty realistic, and how we didn't manage to hit anything would have been due to the high level of fear and adrenaline. Shooting under extreme pressure would make you're shooting far less accurate than planned and prepared shots at the range, imho.

[^Th0mas^ 10]

Yup, this is the exact same impression I have of the ARMA2 AI.

Good read. Maybe the developers take notice of this post and do something about it.

[mr reality 0]

Its all starting to make sense now, why all we got was faction vids from the devs, and no real vids showing off this all new AI.

Reading these snippets of info has saddened me somewhat, but i suppose i'll just enjoy ArmA2 for what it has to offer,. and not what i had really hoped for.

ArmA2 is looking to be a big dissapointment, as it's looking more like Armed assault everyday with not much in the way of major changes. I'm still more than a little pissed at the crappy AI driving and the shitty command speech.

[Potatomasher 0]

Its all starting to make sense now, why all we got was faction vids from the devs, and no real vids showing off this all new AI.

Reading these snippets of info has saddened me somewhat, but i suppose i'll just enjoy ArmA2 for what it has to offer,. and not what i had really hoped for.

ArmA2 is looking to be a big dissapointment, as it's looking more like Armed assault everyday with not much in the way of major changes. I'm still more than a little pissed at the crappy AI driving and the shitty command speech.

I suggest you try Arma 2 yourself before locking up your opinion. You may be positively surprised.

Many people said that much hasn't changed since Arma 1 and they are right to some point. The core of the game is obviously identical to Arma, but still something in this game got me hooked much more than it did with Arma.

I don't know is it the terrain, new animations, better sounds or better gfx or what but i think this is the proper sequel to OFP, not Arma 1. Sure there are lot's of AI bugs and performance problems like every BIS games have had in the start. But with patches i think i'll be playing this game a long time(like OFP). I ditched arma pretty soon only to install it again to try CWR mod recently. This ones better i can feel it.

[fabrizio_t 58]

Its all starting to make sense now, why all we got was faction vids from the devs, and no real vids showing off this all new AI.

Reading these snippets of info has saddened me somewhat, but i suppose i'll just enjoy ArmA2 for what it has to offer,. and not what i had really hoped for.

ArmA2 is looking to be a big dissapointment, as it's looking more like Armed assault everyday with not much in the way of major changes. I'm still more than a little pissed at the crappy AI driving and the shitty command speech.

Well, to my opinion ARMA2 resembles more OFP than ARMA. That's why i enjoy it. Never really liked very much ArmA, i think it was lacking come "soul".

[POTS 0]

How long were the firefights people had so far? For me the fights are still rather short and in concept very similar to the way we fought in OFP and ArmA.

The new AI shows some impressive abilities compared to ArmA but it seems mostly about moving close to cover, not to actually fight from cover. Kneeling next to a corner or tree is nice for looks but it provides as much cover as kneeling on open ground. As such it is my impression after playing Arma 2 for two weeks pretty intensively that combat is not fundamentally different than in ArmA.

There is AI that leans around corners or shoots over small walls but it happens not that often. Of course it has also to do with the fact that, save for urban terrain, there is not that much cover available in the world (as trees are no cover). The AI seems generally to be very reluctant to fight from good positions should they find one, moving further quite quickly. Also I do not have the impression that suppressive fire works for the AI on the receiving end. I do not say it isn’t in the engine but while playing I could not yet see it making any noticeable difference.

The concept of fighting from sheltered positions, exposing only as little as possible when firing and to quickly duck back to full cover when under fire, still seems to be missing from the AI. I hoped for longer firefights from clear positions in Arma 2 but my impression is that combat still plays pretty much the same as previously.

The reason why firefights can be short is because of the overdone zoom.

[fabrizio_t 58]

Truth explained about squaddies knowing your position when you kill any AI group members

Many people already report about AI groups becoming suddenly aware of player position when he kills any group members, with them unaware of his presence and far from their leader / squaddies.

Well, that's not entirely true. You can kill any AI units without having their own groups being informed of your presence/position, but there are two prerequisites for that to happen:

1) camerades of the victim should be unaware of you and far enough, not to hear your shots;

2) you should kill your victim with a single shot;

If you won't kill with a single shot, your victim will automatically gain some "knowledge degree" of your position and that will be INSTANTLY broadcasted to their mates, anywhere they are.

Technically that "knowledge degree" is stored in the "knowasabout" variable: the bare minimum value you can expect for "knowasabout" if you miss a single-shot-kill yet you were concealed / not visible / not audible is "knowasabout"=1.5. That's a pretty high value that allow victim's groupies at least to actively look for you.

So yes, we have a major issue here, as it is simply a consequence of 2 precise problems:

1) The MINIMUM knowledge (about the shooter) that an AI unit gains when kiled, is TOO HIGH ("knowasabout" = 1.5), since it does not take in account if the shooter was concealed, audible or visible.

2) The "knowasabout" value is INSTANTLY broadcasted through groups, so when a unit is hit by any shot all group members will share the same knowledge about the shooter.

---------- Post added at 01:01 PM ---------- Previous post was at 12:47 PM ----------

EDIT: sorry, i have to correct myself: after further testing it seems that even if you kill your victim with a single lethal shot his group will gain at least a knowsabout value = 1.5 of you in a matter of few seconds. This sadly happens also if the victim group is > 1 mile far. :(

Test mission:

http://rapidshare.com/files/244074758/check_AI_kill_awareness.Chernarus.zip.html

(sorry, just 10 downloads allowed)

Edited June 13, 2009 by fabrizioT

[mr_centipede 31]

I think knowsabout is not really the location of the shooter, but rather that they know there is an enemy nearby so they will search more vigorously.

One thing I agree though, their network of info sharing is too quick. I think there should be some lag let's say 10-30secs?

[fabrizio_t 58]

I think knowsabout is not really the location of the shooter, but rather that they know there is an enemy nearby so they will search more vigorously.

One thing I agree though, their network of info sharing is too quick. I think there should be some lag let's say 10-30secs?

To my knowledge a "Knowsabout" value = 4 means that a unit knows also the exact position of another unit (usually this value apply to threats that are in the line of sight of a unit).

A "Knowsabout" value of 1.5 will make any unit move towards the general location of the threat, so i think that "Knowsabout" ALSO conveys some kind of information about the location of the threat.

EDIT: I have just noticed in the BI Wiki that the issue i depicted here was confirmed in OFP (!).

see: http://community.bistudio.com/wiki/knowsAbout

Also... As soon as An AI unit gets hit with a bullet his knowsAbout level for the unit that shot him will instantly jump to 1.5, even if the shooter is 1000m away. Vice versa is also true. If An AI soldier's knowsAbout level for an enemy is 0.7 and the soldier fires and hits the enemy, his knowsAbout level about the enemy will instantly jump to 1.5. ...

My points are:

1) I have really no problem with the victim gaining knowledge about the shooter, but the victim should be STILL ALIVE then (so single lethal shots should not raise knowasbout).

2) Also the minimum gained knowasbout value of 1.5 is TOO HIGH and does not take in accout shooter concealement, visibility and audibility.

3) Finally the instant broadcasting of knowsabout through group units makes no sense.

Edited June 13, 2009 by fabrizioT

[walker 0]

Hi all

Nice to see proper experiments being done! Now perhaps people will see why it is important.

On the matter of knowsabout it is the accuracy they have on your position probably in terms of angle, in 3 dimensions not just 1.

The ideal for me would be that once the 2 D arc gets to say 10 or 15 degrees the enemy starts to methodicly or speculativley supress that arc, sometimes from the centre out sometimes from edge to edge.

That a knowsabout value of 1.5 causes the enemy to run toward the source, is reasonable but it would be better if they ran toward anywhere on that 45 degree arc rather than directly at the source. And then as their knows about values increase they should repath and home in.

The amount of time the AI takes to communicate the enemy position should be how long a player takes, close in it should be 1 to 2 seconds eg. saying contact right or contact left takes about that long

There was something like this in an old ArmA patch where they used the completion of the phrase by the AI commander but that caused bugs with translation times. I think it would be better that it is hard coded on distance.

1) 0m to 100m it you'r saying Contact Right, close. So it takes 1 to 2 seconds.

2) 100m to 300m you'r saying: contact right, 250m, 110 degrees, sniper. So about 5 seconds.

3) Beyond 300m you'r saying: contact right, 250m, 110 degrees, reference electricity pylon, left 30m, up 20m, in that bush, sniper. So about 10 to 15 seconds.

Kind Regards walker

Edited June 13, 2009 by walker

[fabrizio_t 58]

Hi all

That a knowsabout value of 1.5 causes the enemy to run toward the source, is reasonable but it would be better if they ran toward anywhere on that arc rather than directly at the source.

That's what happens.

What's not reasonable is all AI units in a group get knowsabout = 1.5 when you fired a single lethal shot from 1000m away wearing a ghillie suit prone behind a three ...

That minimum knowsabout value for each units should at least take in account:

1) how far the shooter was.

2) how well he was concealed.

3) which class the shooter was (Rifleman? Sniper?).

4) how loud was the shot (= which weapon was used).

5) minimum knowsabout should be also constrained to the skill for each unit of the group, unless the shooter is explicitly revealed by the group leader

Edited June 13, 2009 by fabrizioT

[KeyCat 131]

@ fabrizioT: The enemy detection routines are not that simple and not only related to the knowsAbout value. Exactly how it works internally I don't know but for these types of test I think it's better to use nearTargets command to see AI's percieved position of you.

However, after reading this and various AI threads it sure looks very similar to some issues already discovered in ArmA (again, see this thread) but I will have to wait until I tested ArmA II myself to form my opinion. I'm also trying to remind myself that the German 1.01 release is something like a "pre release" and I'm sure we will see improvements in this (and other) areas!

PS: If you look inside my testmission (linked in above thread, page 4) you find a simple ArmA script you maybe can adapt to ArmA II.

EDIT: One more thing..... when doing the tests please note your actual skill settings (both in x.ArmAProfile and skill slider in editor) along with the testmission if others wan't to reproduce it.

/KC

Edited June 13, 2009 by KeyCat

[fabrizio_t 58]

@ fabrizioT: The enemy detection routines are not that simple and not only related to the knowsAbout value. Exactly how it works internally I don't know but for these types of test I think it's better to use nearTargets command to see AI's percieved position of you.

However, after reading this and various AI threads it sure looks very similar to some issues already discovered in ArmA (again, see this thread) but I will wait until I tested ArmA II myself to form my opinion.

If you look inside my testmission (linked in above thread, page 4) you find a simple ArmA script you maybe can adapt to ArmA II.

EDIT: One more thing..... when doing the tests please note your actuall skill settings (both in x.ArmAProfile and skill slider in editor).

/KC

Ok, here you go:

if i kill with a single shot an unaware enemy near me (in a church courtyard) his group leader, that is 1400m. far and completely out of line-of-sight will perceive my position with an average error (radius) of about 38-45m. That was checked through the position given by NEARTARGETS. By the way, i fired a MP5SD and was wearing a ghillie suit.

Doing the math i think that AI leader is looking for my position in a sector that is 3-4 degrees wide. That obviously without having seen me.

May i say that's a bit scary?

To me that simply confirms the issue that was portrayed by looking into the "knowsabout" value;

With these settings i think there is no point in sniping, nor ambushing, nor doing saboteur activities with explosives ... this is a MAJOR issue.

About skill: i always test cases with maxed skill for all units in editor, just forgot about giving out that info, sorry.

In "armaProfile" i have both "precision" and "skill" values lowered to 0.8. So things may eventually be worse than how they're depicted here.

Edited June 13, 2009 by fabrizioT

[KeyCat 131]

...

About skill: i always test cases with maxed skill for all units, just forgot about giving out that info, sorry.

Maybe thats where you should start? I mean, why complain of to good AI if we are able to tone it down ourself? Maybe the "fix" is right there infront of you? I'm not saying you guys are totally wrong in your observations but IMHO AI at max should be very very hard to beat!

I'm sure that current ArmA II version have it's share of issues and needs tweaking/fixes but I suspect that some of these AI issues may come from the use of to high AI settings?

/KC

Edited June 13, 2009 by KeyCat

[Daniel 0]

AI settings (in the editor at least) should not affect whether the AI is "cheating" or not. If it does, it's pretty useless. In my opinion, it should set the professionalism of the individual unit, e.g. its accuracy, its morale, its weapons use, its tactics.

So a unit set with low skill should spray more with an AK, be more prepared to flee, be less likely to use complex tactics, and be generally less accurate.

High skill setting should give a unit slightly better marksmanship (but not beyond that of the average professional soldier), correct use of weapons (so single shot or 3-5 rounds bursts depending on weapon, etc), and the ability to perform manouvers such as bounding overwatch and peeling.

Actual AI settings (such as their ability to pinpoint unrealistically after being engaged before sniping back unrealistically, as has been suggested here), should be addressed by BIS in the engine. This should be unrelated to the skill slider in the editor.

Don't see any harm in overall AI precision remaining an option in the profile config.

Edited June 13, 2009 by Daniel

[fabrizio_t 58]

Maybe thats where you should start? I mean, why complain of to good AI if we are able to tone it down ourself? Maybe the "fix" is right there infront of you?

I'm sure that current ArmA II version have it's share of issues and needs tweaking/fixes but I suspect that some of these AI issues may come from the use of to high AI settings?

/KC

I see you are not concerned about the issue. Sadly i am.

However let me add that if the enemy group leader (and his mates) are given skill=ZERO in editor, they will track your position with the exact SAME accuracy as with maxed skill.

Now let me please know whether or not should i complain ...

---------- Post added at 02:50 PM ---------- Previous post was at 02:37 PM ----------

Here is a revised test mission that shows also the difference between player position and the enemy-perceived player position.

Just kill the OPFOR unit in front of you and wait a few seconds ...

I measured the error in perceived position like 30 times now, it's in the range 2m - 65m (from 1400m. distance) regardless of the skill of the enemy group units.

So enemy units after you kill a squadmate with a single shot will pinpoint your position from distance and through cover into a sector that is 1-5° wide.

http://rapidshare.com/files/244103465/check_AI_kill_awareness_2.Chernarus.zip.html

Edited June 13, 2009 by fabrizioT

[KeyCat 131]

I see you are not concerned about the issue. Sadly i am.

If you search this forum you will notice just the opposite! I'm actually very concerned about these kind of issues as well as the AI in general. Like many others I spent countless hours in OFP/ArmA to prove wierd AI behavior/cheats.

Again, I'm not saying you are wrong in your observations since I can't test anything myself until I have ArmA II installed on my HD.

What I say is that it it may be a good idea to perform (and document) the tests with different skill settings etc. before jumping to any conclusions.

/KC

[fabrizio_t 58]

If you search this forum you will notice just the opposite! I'm actually very concerned about these kind of issues as well as the AI in general. Like many others I spent countless hours in OFP/ArmA to prove wierd AI behavior/cheats.

Again, I'm not saying you are wrong in your observations since I can't test anything myself until I have ArmA II installed on my HD.

What I say is that it it may be a good idea to perform (and document) the tests with different skill settings etc. before jumping to any conclusions.

/KC

Ok, sorry, my fault, i misunderstood.

So far i've made many tests, i've looked into "knowsabout" and checked "neartargets", i've changed units skills and units classes.

I've changed the test location too.

Finally i provided a test mission.

I hope somebody will be able to check it out and to give feedback about it.

Edited June 13, 2009 by fabrizioT

[walker 0]

Hi all

fabrizioT is doing it the right way. As has Keycat in the past and I am sure will in the future, as soon as like me he gets his copy of ArmA II.

It is only with real repeatable experiments that one can argue for changes in how the AI works.

Anything else is just people letting off gas.

Kind Regards walker

Edited June 13, 2009 by walker

[KeyCat 131]

Ok, sorry, my fault, i misunderstood.

So far i've made many tests, i've looked into "knowsabout" and checked "neartargets", i've changed units skills and units classes.

I've changed the test location too.

Finally i provided a test mission.

I hope somebody will be able to check it out and to give feedback about it.

No worries fabrizioT and I do appreciate your (and others) tests so far and to actually help BIS making a better game/sim we need to provide them with more than just our "opinion" if we can, stuff that they can reproduce in their end and hopefully fix.

As I see it we are all on the "same side" and all/most of us want's the same thing, i.e the best ArmA II experience we can have and I'm sure we will get there with some patience...

/KC

[froggyluv 2136]

@fabrizioT -first off, good test and very interesting stuff.

Ok, tried your mission and yes after shooting 'victim' who is at very close range, the knowabout did jump to 1.5 and they guessed my position to within 45 meters or so. Then I tried shooting him and moving away, and their assumed position of me declined with each step I took away, so perhaps their just assuming your in a close radius of the victim. Next I added a trigger that when 'Victim' goes down, they would 'search&destroy' the area within 50m radius of me, but they never did pinpoint me. Lastly, I exchanged your mp5 for regular sniper rifle, moved "victim" approximately 100 meters NW of his regular location, and took the shot with me 100m to his south -guess what, 0 knowsabout and no reading that the 'group' knew anything about it. 2 times out of ten on the last scenario, the 'groups' knowsabout did rise to 0.1, but it dropped after approximately 20 seconds and no reading of my perceived location.

[fabrizio_t 58]

I see that people that do not have the game may have problems in evaluating whether or not the the test case is reliable, so i'll tell now how my test case is structured.

GOAL

------

- Evaluate whether or not an AI group is able to automatically acquire the approximate position of an emeny who killed an out-of-sight member.

- Eventually evaluate the precision of the AI perceived position of the killer.

MISSION STRUCTURE

----------------------

1) i placed a group of 3 OPFOR units looking NORTH at the top-left margin of the map. Skill was set to ZERO in editor for them.

2) i placed A USMC sniper (player) in a church courtyard, looking NORTH. Distance from OPFOR group was 10.000m. (10Km). Removed all weapons and gave him a MP5SD and ammo.

3) i moved the "victim", one of the previously placed OPFOR units, just 40m. in front of the player, in the same courtyard, looking NORTH (so he could not see the player). I've set his init "Field" to "this disableAI 'move'" in order to have him not moving towards his leader.

4) i created a .sqf script that looped measuring each second the "knowsabout" about player for both OPFOR group leader and the OPFOR "victim".

5) i added some lines to the .sqf in order to show the distance in meters between player position and the position of the player perceived as by the enemy group leader. That distance was given as output for a NEARTARGETS function.

OBSERVATIONS

------------

- Both player and a "victim" are isolated, in cover and concealed against the rest of the group.

- the victim is not aware of the presence of the player at startup

- the squadmates of the victim are 10Km. far, can't see nor hear the player.

AND NOW?

------------

- Player will kill the "victim" with a single shot of his suppressed gun.

- We will check if the far away OPFOR group will gain any knowledge about him and try to measure how accurate that knowledge will be eventually.

RESULTS

----------

- The OPFOR group, that is 10Km far from player and can't see or hear him, with acquire his position (after the kill) with an average error of about 50m.

- Changing the "class" for one or more units of the OPFOR group does not significantly changes the results.

- Maximizing the skill of the units of the OPFOR group does not significantly changes the results.

---------- Post added at 04:32 PM ---------- Previous post was at 03:51 PM ----------

---------- Post added at 04:32 PM ---------- Previous post was at 04:32 PM ----------

@fabrizioT -first off, good test and very interesting stuff.

Ok, tried your mission and yes after shooting 'victim' who is at very close range, the knowabout did jump to 1.5 and they guessed my position to within 45 meters or so. Then I tried shooting him and moving away, and their assumed position of me declined with each step I took away, so perhaps their just assuming your in a close radius of the victim. Next I added a trigger that when 'Victim' goes down, they would 'search&destroy' the area within 50m radius of me, but they never did pinpoint me. Lastly, I exchanged your mp5 for regular sniper rifle, moved "victim" approximately 100 meters NW of his regular location, and took the shot with me 100m to his south -guess what, 0 knowsabout and no reading that the 'group' knew anything about it. 2 times out of ten on the last scenario, the 'groups' knowsabout did rise to 0.1, but it dropped after approximately 20 seconds and no reading of my perceived location.

@froggyluv:

Can't reproduce your results.

If i equip a sniper rifle "knowabout" always grows over 1 after the kill (for the sniper rifle it's 1.4, for other guns 1.5).

About the lack of the perceived distance message it's probably a bug with the poorly refined code that handles the NEARTARGETS output in my main.sqf.

Please try running the fixed test mission linked below (it includes fixed code).

The mission includes a sniper rifle for the player and now the OPFOR group is 10km away.

Be sure that ONLY one WEST unit is present in nthe case you are manipulating the mission, otherwise the .sqf code will fail.

Also wait at least 30 seconds after killing the "victim" in order to see the various messages.

Fixed mission:

http://rapidshare.com/files/244133961/check_AI_kill_awareness_4.Chernarus.zip.html

In case you do not want to download the mission again here is the fixed main.sqf:

Fixed main.sqf code:

_leader = _this select 0;

_shooter = _this select 1;

_victim = _this select 2;

_group = group _leader;

_victim disableAI "move";

_victim setunitpos "UP";

HINT("KILL THE OPFOR UNIT IN FRONT OF YOU. SEE WHAT HAPPENS TO KNOWSABOUT FOR BOTH THE VICTIM AND ITS OWN GROUP (THAT IS 10Km FAR!).");

_t = 0;

while {true} do

{

_tgt = _leader nearTargets 15000;

_pos = 0;

{

if(format["%1", (_x select 2)] == "WEST") then

{

_pos = (_x select 0) distance player;

};

} foreach _tgt;

player globalChat format["%1 - OPFOR Group knows about PLAYER: %2", time, _leader knowsabout _shooter];

player globalChat format["%1 - OPFOR Victim knows about PLAYER: %2", time, _victim knowsabout _shooter];

if(_pos != 0) then

{

player globalChat format["%1 - OPFOR Leader perceived player position with an error of: %2 m.", time, _pos];

};

sleep(1);

_t = _t + 1;

};

Edited June 13, 2009 by fabrizioT

[Sniperwolf572 758]

Just a reminder to test kills via headshot vs bodyshot, from the tests that I've done ages ago in ArmA1, it made a huge difference on the awareness of the nearby enemies to the death of someone nearby. That's where the difference in your tests is probably coming from. Can't remember if they were grouped or ungrouped in the tests tho.

Edited June 13, 2009 by Sniperwolf572