AI Discussion (dev branch)

[Gwynbleidd 28]

There is also a problem where AI is close to a building or a wall, and hears/detects the enemy on the other side, then throws the grenade (in the right direction, spot on actually, the problem is, there is a wall in front of them). It seems to be the same problem as AI hearing you, and trying to shoot you through the wall, but in this case grenade just bounces back to them killing the AI, and injuring his teammates. Either that, or they just don't know there is a wall in front of them.

[cosmic10r 2331]

 

Looking at your video again, this time on a real screen and it's really interesting since the AI reactions differs a lot in your test...

 

Things I notice... you fire 6 shots and are less concealed than I use to be (I hide in the bush to your left), I only fire 1 and max 2 shots and when the squad leader turns in your direction he doesn't spot you (when I test it take 1-2 seconds after he looks in my direction).  You even roll on the ground a couple of times after firing, I lay completely still.

 

A few questions.

 

1. What version are you using in the video, 0.50 or 0.10?

 

2. Would be great if you could check what AI settings are at in your <name>.Arma3Profile is or

   maybe upload it somewhere if you are unsure.

 

Now the question is why are the result is so different? Looking at your HW specifications i see you are using a 6 core 3930k @ 4.2 GHz. Myself use a ancient 2 core E6850 @ 4.22 GHz. I know 2 vs 4-6 shouldn't make any difference and I'm just stabbing in the dark here but maybe it is???

 

Videos are great and next time I will try to make a video showing the result when I test...

 

/KC

 

 

 

 

I'm sure KeyCat wasn't analyzing hardware, he was analyzing AI in devbranch. As he gave his settings and wanted others to try it's assumed you would have the same settings. As the result was so different and the only thing he could see was that you had more cores he was probably trying to find if that was the reason why.

 

 

 

 

1 )  Actually KeyCat was asking about hardware... or are questions about  cpus not hardware anymore... did i miss the memo...

 

 

 

 

 

 

Until you told us you 1 - use a different setting (0.1) which will change the test scenario and make it hard to compare. And 2 - You use an AI addon, specifically made to alter the AI behaviour (ace AI)! Then your test is of no use in testing the dev branch AI.

 

 

 

2) 0.10 is the skill he set in his own mission as a default  for testing so it could be compared...

 

 

So please disable all addons, specially all AI altering addons. Set you settings to what KeyCat use and try the same scenario again. Then KeyCat can compare and we can get some conclusions of the dev branch AI :)

 

 

I guess it wasn't clear that I did that for the last video a few posts ago (apples to apples)  and got terminated... so in this case I  was able to reproduce KeyCats issue.

I think the more interesting ? is what have I done to change that since it doesn't seem like the slider adjust it as much  as one would like...

If I could tweak a bit from the stock to where I have it with  all my mods active it may be quite fetching....

[seba1976 98]

We may be talking about different things here as KeyCat was analyzing hardware and wondering whether the 4 extra thread on my processor were leading to the difference in results in our tests ...

Yes, I stand corrected. Missed the part where he speculated about different proccesors leading to different AI behaviour...

[KeyCat 131]

Thanks for taking the time to make a video Cosmic. The behavior you reproduce is exactly what I mean that is "uber human". In the test you fire one single shot with the silencer on, killing one squad member and one of them (in my tests it's often the squad leader) turns in your direction and pinpoints you as well as communicate your location all within a few seconds. Add to the fact that we are prone in the grass/bush wearing a ghillie suit at ~200 meters distance as well as having the skill set to 0.10 makes the AI over the top, at least in these situations IMO.

 

Even if you take up the best possible position inside that bush with tall grass in front of you (see my previously posted screenshot how it looks like from the AI perspective) which in RL would make a soldier wearing a ghillie suit close to impossible to pinpoint that quick even if he fired a couple shots without silencer. If some of them was facing you and scanning for threats then a big "may be possible" but in the test they are all standing with their backs towards you.

 

Another thing I noticed that seems weird is that if you only wound that soldier with one shot the AI don't act the same and seems more confused and have much harder time pinpointing you..why?? I mean if a squad member is killed with one shot he has no chance to communicate anything but if he is wounded he may have a chance to call out at least some hints about direction, it doesn't make sense. To simplify... most of the time it feels like one squad member killed and one single pixel of you are in LOS = instant knowledge of your position no matter how concealed you are.

 

I think I've mentioned it before but there was actually same/very similar issue where the AI was able to triangulate you thru their "hearing" way back in A1 that was fixed by Suma and it's probably possible to dig up that thread if it can give any hint's, I even have the old repro mission somewhere in my archive :)

 

BTW: I actually tried to make two very short videos the other night but they ended up 1+ GB and that would take ages for me to upload using my thin pipe so screenshots is the best I can offer at the moment (also tried to compress to mpeg but then it was to blurry to see the details).

 

 

I have a hard time believing that the extra threads contribute to a different ai experience. .. Is that even possible?

 

My guess is that it's not really possible and only a stab in the dark since that was the only real difference in our setups I noticed until you said you used an addon.

 

Regarding dev-branch testing and addons. I'm pretty sure the dev's disregard any feedback/videos/screenshots if any 3rd party addons is used and thats understandable. In short.... never use any 3rd party addons if you plan to give feedback and preferable also write down AI settings and exact build number.

 

/KC

[cosmic10r 2331]

It was interesting to see how radical a difference there was in the two tests... I don't exactly remember what I did in the ace ai pbo but I think it helps narrow down where the issue lies by testing using addons vs vanilla. I'm actually kind of shocked that it had that much affect. .. I think the next step for me is to try to tweak it so that there is a distinct difference between the reaction to the silenced shot vs unsilenced...

I agree that the target knowledge is unusually fast using default...

In several of my posts I had a disclaimer on the ace ai. Pbo but I guess it was missed...

The goal has always been that believable ai and as I mentioned testing your mission in default vs my modified ai has given me some ideas on how to get there... :)

[andersson 285]

I wish for BI to get there so we all can benefit.

[rübe 127]

Another thing I noticed that seems weird is that if you only wound that soldier with one shot the AI don't act the same and seems more confused and have much harder time pinpointing you..why?? I mean if a squad member is killed with one shot he has no chance to communicate anything but if he is wounded he may have a chance to call out at least some hints about direction, it doesn't make sense.

Yeah, I also believe that "information sharing"/communication isn't adequately modelled, be it inner-group or even faction-wide communication. Maybe even some nasty bug has sneaked in in this regard. Who knows, inner-group communication might currently just be broken.

 

Question: do the equipped radio (items) in any sort or way affect AI "information sharing"/communication? Like... at all? Would a "guarded by"-waypoint be affected by units not having the radio item?

 

Either way, I see lot's of room for improvements here. AI can move and shoot just fine. The problem arise mostly on a higher level: what are they doing and why? Thus shouldn't there be an easy system allowing mission makers to model means of communication? I'm not talking ACRE here, but an easy way to link up groups or factions even (maybe an additional radio-item would come in handy). You know... like something better than a "guard"-waypoint?

If 5 groups are guarding/patroling some strongpoint/common AOO, we probably wanna have them in constant (radio) contact. We should be able to easily model that in the editor (or by means of scripting - of course). Similarly, and a level higher up, there should be a way to radio for backup/assistance/help. Again: the "guard"-waypoint should be deprecated. Give us something better, where we can define which guys "on guard" are responsible for which groups possibly calling them. Maybe let us define the delay-time from observation (of a threat) to reaction (radio) similar to trigger timeouts (easy stuff for the editor). And make sure that things do not work without equipped radio (maybe introduce another model/item). What if we could link some communication channel (say from group A to B ) to some object (like an antenna) that needs to be alive (and could be destroyed).

 

Given the dynamic nature of ArmA, information sharing/communication should be a driving force in AI decision making (high level). Now sure, one could argue that it's perfectly possible to customly script and model all this right here and now (surely it has been done already, in one way or another). IMHO it should be a part of the engine, seeing how closely related to AI it is. And I'm talking way beyond flanking maneuvers of a single group. What if group A, facing a tank, has no AT, but could check in with HQ, if someone else could maybe assist. Group A would go duck'and'hide (and maybe laser pointing some targets), while CAS (or artillery, or...) is called in (some group B put down on some airport and comm-linked to HQ; maybe even registered for very specific tasks/requests).

 

tl;dr: Let mission designers quickly link groups (2 layers: layer 1 to merge groups in a common AOO, layer 2 to link up with HQ), thereby defining/modelling communication channels (finetuned with radio items or linkes communication objects such as antennas) and have some easy message passing the mission designer could configure (requests/on guard). Then let BIS fellows work on inter-group AI, which currently is not possible/has to be scripted on top of the engine. Sit back and watch impressive AI reactions and interplay! But at the very least: fix inner-group communication. ;)

[inlesco 233]

Had a look

 

Lovely video there, cosmic.

All we need now - a compilation of all AI video tests and a thread with all of them listed and sorted by significance to the gameplay and overall experience.

 

A couple of advices for improving AI behavior:

 

----  When under fire by an unknown entity [with a silenced weapon] / a sniper from far away, AI should:

• depending on their [probably false] guess on the caliber of the wpn, seek cover immediately (not talking about signposts / roadsigns which serve as cover from snipers :D)
• when cover is unavailable / too far away, take the risk to scan the area in all directions immediately. If a soldier was hit in his back, it's obvious from where the bullet is coming. As shown in the video, AI doesn't understand this simple concept of fire direction.

Now, several considerations of current AI behavior:

• AI doesn't scan 360 degrees, they just look 'somewhere'. Sometimes, at a static building or a fence.
• At 3:06, a shot missed and the AI correctly guessed the direction by the angle of the impact. That's quite impressive, IMO.
• Or I'm mistaken and the AI still uses its godly powers whenever pissed :D

Btw, is there any BIS' AI testing framework with all variables comfortably logged / displayed in the UI so we don't have to use the hints?

I mean, come on... It's 2015 already.

[Greenfist 1863]

My detection script that keeps coming up here was initially made only for testing AI's ability see through objects. I'm glad you found other uses for it too. :)

But the color coded symbols didn't really work with more than one unit, as you can see in the videos. Like the red indicator was supposed to mean you're currently in the line-of-sight, but at times they obviously have no idea where you are.

 

So I improved it a bit to take multiple units into account.

Download link and some instructions:

// AI detection test v5
https://www.dropbox.com/s/zs0e86db28h3nwz/AI-detection-test.Altis.zip?dl=0

White icons = AI knows nothing about you.
Blue = AI has seen or heard you, but doesn't know if you're hostile or friendly.
Yellow = AI knows you're an enemy, but haven't seen you yet.
Green = You've been spotted but not currently visible.
Red =  AI sees or hears you.

 

The mission revolves around an AI unit named 'enemy1'. The text box in the bottom corner shows some data about him and the big floating arrow shows the position where he thinks you are.

Actions:
    Look at me: Turns enemy units towards you.
    Teleport enemy: moves enemy to cursor position.
    Teleport player: moves player.
    Reset enemy: enemies forget about you and turn away from you.
    Virtual arsenal: just that.
    Get down/up: changes his stance to prone/standing.
    Splendid cam: it's splendid isn't it.
    Enemy cam: Shows the view from enemy's eyes for a few seconds.
    
Shift + Left-clicking on the map: creates a move waypoint for the enemy
Alt + click: teleport player

To move this script to another map:
    Copy this init.sqf and GF_fncs.sqf to your mission folder.
    Preview.

 

By the way, the AI's field of view is 120 degrees and they turn their head all the time. That's probably why they seem to spot enemies at so unrealistic directions sometimes.

[KeyCat 131]

Lovely video there, cosmic.

All we need now - a compilation of all AI video tests and a thread with all of them listed and sorted by significance to the gameplay and overall experience.

 

Just a reminder!

 

The video in your quote (the first test Cosmic did with my simple test mission) are made using 3rd party addons and that invalidates the test completely.

 

The second video Cosmic made using vanilla A3 is the one that should be used to observe AI reactions in this particular test/situation and as you can see below it's a totally different result...

 

https://www.youtube.com/watch?v=Dj0rDSJLyOs

 

@Cosmic: You should probably remove that video from this thread or at least put in a big fat disclaimer so no more misunderstanding occurs.

 

 

Btw, is there any BIS' AI testing framework with all variables comfortably logged / displayed in the UI so we don't have to use the hints?

I mean, come on... It's 2015 already.

 

Some kind of common AI testing framework from BIS would be totally awesome!

 

Even a "simpler" mission made by someone in the community having a variety of isolated test situations like short/mid/long range engagements, CQB, sneaky stuff, sniper stuff, AI vs armour etc. on different locations on the map so they don't interfere. Additionally three selectable times of the day, morning, midday and night (maybe also some selectable weather) would help but this can also be changed in the Splendid camera.

 

@Greenfist: Thanks for improving the script it's a useful tool, will update my mission.

 

/KC

[seba1976 98]

It was interesting to see how radical a difference there was in the two tests... I don't exactly remember what I did in the ace ai pbo but I think it helps narrow down where the issue lies by testing using addons vs vanilla. I'm actually kind of shocked that it had that much affect. .. I think the next step for me is to try to tweak it so that there is a distinct difference between the reaction to the silenced shot vs unsilenced...

 

ACE must be changing how spottable the units are, from the config of each unit. It's a common technique. Than would explain the difference very easily.

[SilentSpike 84]

Our AI PBO:

 

https://github.com/acemod/ACE3/tree/master/addons/ai

 

If you're testing vanilla behaviour I recommend not running it :)

[cosmic10r 2331]

Just a reminder!

 

The video in your quote (the first test Cosmic did with my simple test mission) are made using 3rd party addons and that invalidates the test completely.

 

The second video Cosmic made using vanilla A3 is the one that should be used to observe AI reactions in this particular test/situation and as you can see below it's a totally different result...

 

@Cosmic: You should probably remove that video from this thread or at least put in a big fat disclaimer so no more misunderstanding occurs.

 

 

Even a "simpler" mission made by someone in the community having a variety of isolated test situations like short/mid/long range engagements, CQB, sneaky stuff, sniper stuff, AI vs armour etc. on different locations on the map so they don't interfere. Additionally three selectable times of the day, morning, midday and night (maybe also some selectable weather) would help but this can also be changed in the Splendid camera.

 

 

/KC

 

 

I moved the video to this thread... sorry for the confusion guys... 

https://forums.bistudio.com/topic/184708-ai-discussion-any-branch/

 

I have been testing it at range for awhile... I did about 7 tests  from 200 m to 700m and they don't seem to gain the target knowledge the same way... (yes... with DEFAULT AI ;) )

The odd thing was... and i'm not totally sure... but at about 200m it seemed like if I had a building between me and team leader... and I shot a grouped unit... the targetKnowledge didn't suddenly switch to true about 5 seconds later. It stayed false. I was also lower than they were on the terrain.

 

You will laugh tho KeyCat... I reactivated the mods and still had terminator ai and couldnt figure out why. Couldn't go back to the first video settings... thought it was a ghost in the machine... think I have it now but there was a few head slapping moments lol...

 

In any case I will post most videos in the new thread to try to keep this a bit cleaner and avoid any confusion...

 

Thanks Greenfist fot the update! Really appreciate that...

[oukej 2911]

Question: do the equipped radio (items) in any sort or way affect AI "information sharing"/communication? Like... at all? Would a "guarded by"-waypoint be affected by units not having the radio item?

No, AI "always has a radio". AI commanding is tied to it, AI can't really exist without an radio. The radio item itself is more about the audio/subs/VON functionality. Modelling an AI commanding without a radio would be too complicated at this point.

[domokun 515]

No, AI "always has a radio". AI commanding is tied to it, AI can't really exist without an radio. The radio item itself is more about the audio/subs/VON functionality. Modelling an AI commanding without a radio would be too complicated at this point.

This is very significant. Thank you for confirming what I'd long suspected. In fact, I'd suggest that it be pointed out via an in-game tutorial "Listen men, unless proven otherwise, assume everyone of your enemies has a radio and knows how to use it! This means..." As I pointed out in another thread, this certainly isn't the case in the real-world, with the exception of professional armies. Even then, it would be great if players could receive some kind of visual cues (animation) that a combatant is using a radio (tilted head in the case of a headset) or even pulling out a unit (if a handset). Can you imagine the tension in trying to neutralise an combatant who's under fire before he manages to transmit? This could even be extended to PRC (Portable Radio Communication). Usually there's only 1 PRC per platoon and so the tactical importance is paramount, as most PRR (Personal Role Radio) have range limited to 500m. Communications beyond that are impossible other than via the platoon's PRC (manpack's range 3-6km). A manpack PRC is usually large, fairly heavy (4-8 kg) and complicated so it is assigned to a designated radio operator. Taking out the radio operator or better still the manpack would add an interesting element to any tactical decision-making. The manpack would probably then radio a vehicular unit which would have a much larger and stronger PRC (100W instead of 20W). The vehicular PRC is key to calling in support in the form of CAS and/or artillery. The chain of command and latency in these communications seems to be non-existent in Arma which is a shame. While I understand that adding animations and/or introducing AI command without radio would be difficult (although it would be brilliant for low-tech insurgents and/or historical missions), I would suggest that introducing a certain logic to radio communications (PRR > manpack PRC > vehicule PRC > FOB > headquarters) might not be so hard. Maybe worth a proof of concept? At the very least, the speed to information flow could be slowed at platoon level, then out to neighbouring units (depending on range). What do others think?

[seba1976 98]

@domokun: Either ACRE is wrong, or you've got your numbers wrong.

http://gitlab.idi-systems.com/idi-systems/acre2-public/wikis/basic-concepts

[rübe 127]

No, AI "always has a radio". AI commanding is tied to it, AI can't really exist without an radio.

Ah, yeah. I can see how AI commanding (and scripting) get's much more complicated, once communication channels would be modelled and required for AI commanding.

Then again, that only applies to a group (inner-group communication). Modelling inter-group or group-hq communication, however, would be still possible without that "always has a radio" premise. And maybe inner-group communication could at least be adjusted with some parameters/functions (communication delay, ...).

So maybe not for ArmA3, but I hope you guys will think about communication channels while planning the next iteration. IMHO there is a lot of potential for better, more believable AI and meaningful dynamics, easily and quickly setup in the editor (or scripted) just by defining some communication-channels (granted, now that I think about it, this sounds a lot like "high command" for AI).

[seba1976 98]

By definition, you should not simulate things that no one will be able to observe...

[domokun 515]

@domokun: Either ACRE is wrong, or you've got your numbers wrong.

http://gitlab.idi-systems.com/idi-systems/acre2-public/wikis/basic-concepts

Thanks for the correction but my numbers weren't that far off:

Tactical transceiver, e.g. PRC-343 do have a range of about 500m

PRC, e.g. PRC-152 have a longer range of about 5km

PRR, e.g. PRC-117 have a longer ranger of 20+ km

But I think my point remains valid. Where in Arma are the limits of these radio ranges simulated? I'm not asking for ACRE-level attenuation. But could we at least have a few basic rules? e.g.

1) AI can only call support from units within 5km radius (3km if in urban environment) unless his platoon equipped with PRR

2) If the range operator is killed, the PRR is silent until another AI runs over and uses it (maybe impossible for poor-quality units)

3) The transmission of information and decision-making should take more time (authorisation often required by section commanders)

 

 

By definition, you should not simulate things that no one will be able to observe...

 

By that token we should not be simulating suppression or exhaustion? So while I agree that modelling animations of AI using PRC is maybe unnecessary, I think that modelling their ability to contact other AI is not only feasible but relatively easy (IIRC this was simply done using the Group Link 4 mod in Arma 2). Equally I think that it would be good to see AI using their platoon's PRR.

[gammadust 12]

By definition, you should not simulate things that no one will be able to observe...

 

could you develop further that view?

[andersson 285]

Don't waste cpu on things that you don't need to simulate if the observed result is the same?

[gammadust 12]

That is a big "if". I just think that something being "observed" or not is not sufficient to determine if some system should be modeled or not. Enemy communication is a good example, you won't see or hear them communicate, yet it is quite debatable if a model ignoring communication will present a believable end effect. ALIVE for instance will, for processing budget issues, suspend simulation of AI units, caching them when out of sight of the player(s), still employs a custom outcome simulation of opposing AI encounters likely to happen. If the cpu cycles spent in this simpler simulation are worthy is another issue altogether.

 

In simulation terms, out of observation is not by definition and necessarily unmodeled.

[seba1976 98]

Radio operators are key elements, yes, and thus they are positioned, and move, different than other units in their squad. Not quite compatible with the typical headless chicken behaviour we have in the current implementation. Units are not aware of their roles, and any importance put on those, would bring only more awkward situations, without a complete overhaul, which is out of the question as I understand it.

[domokun 515]

I realise that adding a radio operator role in a unit, while preferable, requires a lot of work. So maybe we can push that down the list into the "Nice to have" category. But my remarks are not limited to this, I've also pointed out that the speed and range of AI communication is equally, if not more important. So I'd advocate moving both of these features into the "Should have" category. Also because I think that they would be easier to implement. So let's go for the low-hanging fruit first in our quest for more believable AI.

[seba1976 98]

Share of information between different grups, over the radio or otherwise, is not implemented, as far as I can tell. If some sort of sharing has been implemented, is has a very limited radius (so limited that you can't separate it from hearing gunshots, or generally detecting the enemy units with their own sensors, to do some proper testing). Groups 1 km away from each other won't share any information at all as far as I know. Communication is indeed happening, no matter what, between units of the same group. I believe that's what the dev was reffering to a couple posts back.

 

If a group has a GUARD waypoint, and only in that case, they will move, nonetheless, to the point where units of the same side are being engaged.

 

AFAIK, that's all the intercommunication that exits in the current implementation.