AI Discussion (dev branch)

[Alwarren 2767]

11 hours ago, oukej said:

People have often resorted to low skills in order to make AI less accurate. Sometimes (unknowingly) breaking the AI on lower difficulty presets (or as our QA puts it - making the AI brain-dead ;)) or in different aspects.

This might not have been an issue for missions tailored for a certain difficulty and played privately only with such settings. But anything public was subject to difficulty scaling and parts of AI broken by it.

Moreover some skills are used for several AI aspects at once.
The issue has been troubling us for some time. Our concerns are no less than yours. On the other hand the change opens the gate to a better use of weapon AI properties (aiming), more useful scaling of AI skill through difficulty and overall better and easier balancing.

 

 

Is there any way to give a rough estimate on how the skill setting has changed? It would be beneficial to know so we don't have to start again at zero.

[danil-ch 165]

13 hours ago, oukej said:

:/

Oukej, can you please check this issue https://feedback.bistudio.com/T122157? Both repromissions included.

 

AI and grenades

They seem do throw hand grenades but only if they're out of ammo or they consider it to be cost-efficient to do so. As for GL, they're used them against infantry but extremely rare. Back in 1.00 (and year or two after) situation was much better.

 

Small battle (49 grenadiers)

 

[Vasily.B 529]

@oukej Thanks for the reply. 
I got some idea how to fix rearming right here :
https://feedback.bistudio.com/T119623
As for braking issue, there is just runaway needed and one tank - in which You are the commander, speed up order to stop
https://feedback.bistudio.com/T126349

As for the rest, the issue with tanks heading into forests were mostly on CUP, thanks for answear again, you cleared everything for me.

[krycek 349]

13 hours ago, oukej said:

Jumping on the bait. We are independent. We are a relatively small team. Gamedev isn't a gold mine. Funds don't always multiply dev power. Growth is only good when it's sustainable. Good devs are a rare breed.

When I said indie devs I meant 5 dudes in a rented apartment, not independent. I realize you aren't Ubisoft or EA, BI is more like CD Projekt or Egosoft so it's something in between giants and indie devs. 

If as you say you lack manpower in this case an AI programmer then why you didn't scrapped this system and kept using it for 16 years and 4 games of "well it kinda....works" AI driving?

Like I said earlier it worked in OFP and A1, but with A2 and now in A3 it's becoming a big deja vu for this part of AI.

 

Unlike some people that think only infantry should be the focus and the rest background props, I doubt BI had in mind Napoleon Wars line infantry sim when OFP and Arma series was made. Heck, even your A2+expansion is called A2 Combined Ops.

And let's face it, what the infantry AI does now and their pathfinding it's light years ahead of A2. I had a scenario where an enemy squad had a seek&destroy order and they cleared a house a few streets back where a few of my targets where(and those guys weren't near windows).Compared with A2 it's like the infantry went to university.

 

 For me a convoy or a tank platoon arriving at their objective without one of them being stuck in a big tree or somewhere, flip up or crashing into each other would be a huge improvement.

Even making a temporary workaround like disabling the big bad tree and the stuck vehicle physics while you're 800m-1km away from it will be better than current situation.

Or if that workaround would break things even more then at least make them back up from the tree or house like they do when crashing into each other.

In the vid I posted a few pages back with the tank stuck in the tree if I'm looking at his tracks it keeps going forward so it doesn't even detect the obstacle.

 

Not having any AI driving news it made me think that not only it took a second place, but it took a backseat since 2016. Let alone the fact that we didn't had any AI news since spring.

@oukej your post in Dev Branch AI log is one of the best I read since 2016 seeing things are moving again.

[froggyluv 2136]

 

 While I appreciate the more standardized subSkill results as well as the' low skill too much dispersion' (both help me on my project) -the tone and content of the dev post is another blow to anyone holding out for new AI behaviour -ANY new behaviours. Since AI learned to crawl under fences in Arm 2 we have seen an end to absolutely anything interesting in that area. Sure they are more streamlined in terms of moving thru towns, over small bumps and rocks in the roads etc.. but this is pretty depressing. WIth all of these special teams recruited and built for various projects Visual/Audio/Islands/Jets/Tank Ballistics -I just dont buy that NO ONE can help with the AI - they're are just too deep in the system.

Quote

 AI is predominantly core/engine stuff. That limits external cooperation.

 

I just find this a covenient company workaround. Surely there are things an experienced AI designer or team could do to work with expanding the AI withing the confines of the existing legacy code. But yet again, the focus decided in company roadmap meetings is to improve and push the MP and hopefully the pesky SP will go away or better -be converted! Youre telling me you cant give us a few new animations as reactions? A few more banter lines? The ability to hop a fence instead of crawl it? Some basic self preservation and battle analysis? To write a simple routine to check all Building Positions? All 'too deep' in the code?

 

 Guess im fooling myself. The signs were all there for a long time yet with the promise and glimmer of a new DLC Tac-Ops - seriously? Using "TAC" anything for tactically starved singleplayers, telling us its for the tac minded realistic SP situations -what do you expect us to think? You can tell us its just too hard to go there all you want but all someone has to do is load up C2 Command to see what is possible just thru scripting - this is where Arma should be on its own AI development-wise at this point but yet again we're left in the dust..

[oukej 2911]

1 hour ago, froggyluv said:

 

 While I appreciate the more standardized subSkill results as well as the' low skill too much dispersion' (both help me on my project) -the tone and content of the dev post is another blow to anyone holding out for new AI behaviour -ANY new behaviours. Since AI learned to crawl under fences in Arm 2 we have seen an end to absolutely anything interesting in that area. Sure they are more streamlined in terms of moving thru towns, over small bumps and rocks in the roads etc.. but this is pretty depressing. WIth all of these special teams recruited and built for various projects Visual/Audio/Islands/Jets/Tank Ballistics -I just dont buy that NO ONE can help with the AI - they're are just too deep in the system.

 

I just find this a covenient company workaround. Surely there are things an experienced AI designer or team could do to work with expanding the AI withing the confines of the existing legacy code. But yet again, the focus decided in company roadmap meetings is to improve and push the MP and hopefully the pesky SP will go away or better -be converted! Youre telling me you cant give us a few new animations as reactions? A few more banter lines? The ability to hop a fence instead of crawl it? Some basic self preservation and battle analysis? To write a simple routine to check all Building Positions? All 'too deep' in the code?

 

 Guess im fooling myself. The signs were all there for a long time yet with the promise and glimmer of a new DLC Tac-Ops - seriously? Using "TAC" anything for tactically starved singleplayers, telling us its for the tac minded realistic SP situations -what do you expect us to think? You can tell us its just too hard to go there all you want but all someone has to do is load up C2 Command to see what is possible just thru scripting - this is where Arma should be on its own AI development-wise at this point but yet again we're left in the dust..

I sorry to put it this way but your post is what easily drives devs away from following threads. Not because it's criticizing our work - we love criticism that helps us improve the game.
But having to explain and defend against accusations is painfully tiring. At best it only helps another bug survive another day. If you want to make your dreams a (virtual) reality -> join us https://www.bistudio.com/company/careers

Please try to stick in this thread to the development branch and actual technical AI issues at hand. Please strictly refrain from ranting about roadmap, vision, CoDification, NWO and other hypotheses.

[Grumpy Old Man 3546]

Downloading devbranch now, 7.4gb, curious about the recent tweaks.

Gonna reupload the AIprecision video if there's any changes regarding that slider in the difficulty menu.

Any chance to get a conversion table for the new values or is that off the table (heh)?

 

While the post from @froggyluv isn't the most friendly one, that dude's basically part of the inventory around here.

His post shows a lot of passion for a video game that is pretty much standing all alone in its own niche

and has undergone changes that might make one believe this niche is getting too small for that game, or rather what that game has become.

 

Announcing a change on the devbranch that will break basically all missions that somewhat depend on a tweaked AI difficulty (basically every mission using AI units) with almost 4 years into the games life cycle had to result in a post just like froggyluv did.

 

I still believe arma 2 to be the pinnacle of the series, having a plethora of assets to choose from, civilians (FEMALES ) combined with the alice module, that makes civilians have conversations on the streets,

motorcycles/bicycles which are perfect assets for stealth/black ops/recon kind of missions, the battlefield clearance module that makes you carry injured comrades towards safety, getting heavily wounded on the legs forces you to crawl,

first aid modules that require you to have a medic take care of the injured.

This is rough.

This is war.

 

In arma 3 every class comes with a first aid kit. Get hurt? Just use one of the 20 first aid kits you looted from various corpses.

Getting shot to the legs? No worries, you can still run.

No more civilian modules/females that add a sense of authenticity to missions.

NVGs are either too bright or too dark, they were perfectly fine in A2.

 

Arma 2 was 4 years old when Arma 3 came out.

Now Arma 3 is almost 4 years old and pales in comparison even with 5 noteworthy DLCs (Marksmen, Helicopters, Apex, Jets, Zeus) and all the free stuff (diving, anyone doing regular diving missions, no?, awesome as hell stances, FFV, Slingload, weapon resting, vehicle transport) that has been added.

There are no civilian assets other than the male civilians, few cars and a jetski as of now, this really limits mission making for basically any gameplay type besides pure pvp servers.

The diversity amongst military assets is a bare minimum, one mrap, one tank and 2 apc/ifvs per faction, you don't need diversity for pvp type of gamemodes.

 

I enjoy arma 3 and think it's a great game, but it just isn't anywhere near the point that arma 2 was pointing towards.

 

Cheers

[R0adki11 3949]

16 minutes ago, Grumpy Old Man said:

 

I still believe arma 2 to be the pinnacle of the series, having a plethora of assets to choose from, civilians (FEMALES ) combined with the alice module, that makes civilians have conversations on the streets,

motorcycles/bicycles which are perfect assets for stealth/black ops/recon kind of missions, the battlefield clearance module that makes you carry injured comrades towards safety, getting heavily wounded on the legs forces you to crawl,

first aid modules that require you to have a medic take care of the injured.

This is rough.

This is war.

 

In arma 3 every class comes with a first aid kit. Get hurt? Just use one of the 20 first aid kits you looted from various corpses.

Getting shot to the legs? No worries, you can still run.

No more civilian modules/females that add a sense of authenticity to missions.

NVGs are either too bright or too dark, they were perfectly fine in A2.

 

Arma 2 was 4 years old when Arma 3 came out.

Now Arma 3 is almost 4 years old and pales in comparison even with 5 noteworthy DLCs (Marksmen, Helicopters, Apex, Jets, Zeus) and all the free stuff (diving, anyone doing regular diving missions, no?, awesome as hell stances, FFV, Slingload, weapon resting, vehicle transport) that has been added.

There are no civilian assets other than the male civilians, few cars and a jetski as of now, this really limits mission making for basically any gameplay type besides pure pvp servers.

The diversity amongst military assets is a bare minimum, one mrap, one tank and 2 apc/ifvs per faction, you don't need diversity for pvp type of gamemodes.

 

I enjoy arma 3 and think it's a great game, but it just isn't anywhere near the point that arma 2 was pointing towards.

 

Cheers

None of this has anything to do with AI thread, please try to stay on topic. Thanks

[oukej 2911]

On 8/7/2017 at 4:56 AM, Grumpy Old Man said:

Removing the difficulty scaling was not what I meant, what I meant was that when setting a skill via setSkill that it will be the same for every player, no matter if he's running regular/veteran or elite setting for this specific unit without having the skill value being recalculated depending on difficulty setting or cfgAISkill, only another usage of setSkill should be able to change the specific skill.

It's not really a command to set a units skill if it's not consistent across all difficulties, why use setSkill then in the first place if it doesn't have any impact when being tweaked to simulate untrained troops, when someone playing on another difficulty might not even get hit at all because he's on recruit or custom 0.1/0.1 difficulty, another player on elite might not even experience any difference between such a tweaked "untrained" unit and a regular editor placed unit.

That would mean something like setSkillFinal. I'm afraid this is something that would require many exceptions and would be tricky to handle from another user's PoV (why some AI are affected by difficulty and some aren't...)

 

On 8/7/2017 at 4:56 AM, Grumpy Old Man said:

Unable to hit a target at 15m using the precision difficulty slider? That's only possible using setSkill "aimingAccuracy" as far as I tried.

I don't see any difference using AI precision of 0.05, 0.5 or 1.0 in the difficulty settings, maybe I'm missing something.

This is what I encounter on current stable branch, AI skill in difficulty menu at 0.5, precision at 0.04 (somehow jumps to 0.05 every time) then changing precision to 0.5 and 1.0 with no significant change to the actual firing precision also, the "aimingAccuracy" skill somehow always returns 0.75, if that matters at all:

AimingAccuracy skill or skillFinal? First one should always be the same (it's what you set via slider or script). The second one is after AI Level interpolation - the one actually used in simulation.
 

On 8/7/2017 at 8:02 AM, fn_Quiksilver said:

Tiny tweak which would please me greatly (and the players on our server) is to enable Treating/Healing of Incapacitated/Unconscious units. It should be little more than a 1-liner tweak. I left some notes about this in discord convo, and here is the FT for it:

 

https://feedback.bistudio.com/T126026

Revive / incapacitation is currently a scripted/modular system with a specific MP use. Full sandbox support is definitely desired, but atm we've got no ETA. Not anytime soon.
 

On 8/7/2017 at 0:10 PM, Alwarren said:

Is there any way to give a rough estimate on how the skill setting has changed? It would be beneficial to know so we don't have to start again at zero.

The best example is probably the aiDispersionCoef from fire mode config of a weapon. Let's have a weapon with a firemode with aiDispersionCoef 10 (ignoring the Y/X atm). But the actual coefficient that's used to increase bullet spread depends on the skill. The following table (still using that  aiDispersionCoef 10) could better show the difference between then old and new state.

 

AI's aimingAccuracy 	0.1 	0.2 	0.3 	0.4 	0.5 	0.6 	0.7 	0.8 	0.9 	1
resulting coef. (old) 	23.5 	11 	6.8 	4.75 	3.5 	2.67 	2.07 	1.63 	1.28 	1
resulting coef. (new) 	9.1 	8.2 	7.3 	6.4 	5.5 	4.6 	3.7 	2.8 	1.9 	1

You can see the "most interesting" differences were happening ~ between 0.2-0.4 AI aimingAccuracy. Now we are talking about "final" skill - after taking the difficulty/AILevel into account.
To put it in a better perspective - on "Recruit", with AI Level Precision @ 0.2 - an AI with raw aimingAccuracy 0.38 actually translates to 0.194 final aimingAccuracy. And with raw aimingAccuracy 0.2 (still possible via editor) = 0.01 final aimingAccuracy. You can already imagine the resulting dispersion at that value using the old interpolation ;) Broken AI.

 

23 hours ago, danil-ch said:

Oukej, can you please check this issue https://feedback.bistudio.com/T122157? Both repromissions included.

 

AI and grenades

They seem do throw hand grenades but only if they're out of ammo or they consider it to be cost-efficient to do so. As for GL, they're used them against infantry but extremely rare. Back in 1.00 (and year or two after) situation was much better.

AI tends to throw a grenade at a target that the AI doesn't directly see, but it's able to estimate its position and knows it's an enemy.
As for UGL - the AI should be more prone to fire them on condensed groups of soldiers or vehicles - if the probable outcome has a higher value than firing a bullet. But there might be some tweaking necessary.

[Alwarren 2767]

53 minutes ago, oukej said:

Broken AI.

 

 

Thank you for the explanation. I can see how that is a problem, and now I understand why there is such a huge drop-off and almost no change at the higher AI skill range. With the new scaling this almost becomes a linear progression.

Any value between 0.1 and 0.2 with the old scale would make a wild difference. It indeed looks like this is a sensible thing to do.

[Greenfist 1863]

This might explain my test from while back:

Skill in question was "General" but others seemed to scale in similar fashion. More linear future sounds good. :)

[kremator 1065]

On 07/08/2017 at 0:36 PM, krycek said:

 For me a convoy or a tank platoon arriving at their objective without one of them being stuck in a big tree or somewhere, flip up or crashing into each other would be a huge improvement.

Even making a temporary workaround like disabling the big bad tree and the stuck vehicle physics while you're 800m-1km away from it will be better than current situation.

Or if that workaround would break things even more then at least make them back up from the tree or house like they do when crashing into each other.

In the vid I posted a few pages back with the tank stuck in the tree if I'm looking at his tracks it keeps going forward so it doesn't even detect the obstacle.

All of this !  Really makes you think that drivers are retarded.

[danil-ch 165]

5 hours ago, oukej said:

AI tends to throw a grenade at a target that the AI doesn't directly see, but it's able to estimate its position and knows it's an enemy.

It happens so rarely. https://www.reddit.com/r/arma/comments/63swoc/ai_and_grenades/ 

 

5 hours ago, oukej said:

As for UGL - the AI should be more prone to fire them on condensed groups of soldiers or vehicles - if the probable outcome has a higher value than firing a bullet. But there might be some tweaking necessary.

They were very aggressive in alpha and extremely passive now. Hopefully some middle ground will be found.

 

[krycek 349]

I thought of something else for stuck vehicles. If the AI has problems recognizing it's stuck in a house/tree then why not adding a timer and a check for it's current position until it can initialize backing up?

For example "for 15-20 seconds it seems I keep going forward/accelerating towards my position yet I'm still in the same place,bzzzt,time to back up". I assume they have better detection for other AI vehicles when backing up unlike scenery.

 

The same for the flip up or launching into air vehicles glitch. If AI notices he suddenly lost gravity or ended with wheels/tracks belly up a reset timer could be a decent workaround.

I assume they're aware when driving over a surface because otherwise they would dive into the ocean and so on so when something odd happens like his vehicle/tank goes belly up he should know "well this ain't right, time to reset".

[old_painless 182]

Some mod has a feature to move vehicle 5m away if stuck, to "un-stuck". Can't remember which one but a simple solution nevertheless

[lex__1 422]

I have often noticed the following situation when I play on the CTI server. In these missions there are many independent AI. Often AI, in cities, show excellent driving and shooting skills. They drive so fast and dexterously in narrow places (between houses), they are not stopped by obstacles in the form of fences. But at the same time, my AI can stand somewhere near the curb of the road, or at the fork, near the bridge and do not understand - how to overcome them.
But sometimes something happens, independent AIs start to behave very stupid, start to shoot badly, or get a bad reaction. Sometimes there is such a problem, you can go to a group of AI (or any machine with the AI), and they do not react to your presence, you can shoot them one by one (slowly).
Has the intellect disconnected or am I reborn and do not pose a threat to them? In the next city everything is fine, or can happen again in any situation.
I know many places on the mission map where my AI is stuck, with a probability of 100%, but in the AI editor do not get stuck in these places.
I could never reproduce in the editor, any similar problem, to reveal any dependence on the premise of the problem.
This leaves me with an ambiguous view of AI's intelligence problem, or is AI's intelligence code has properties, sometimes stuck in a networked game.

 

 

[fn_Quiksilver 1636]

15 hours ago, oukej said:

Revive / incapacitation is currently a scripted/modular system with a specific MP use. Full sandbox support is definitely desired, but atm we've got no ETA. Not anytime soon.

 

thanks!

 

we are just after the Treat option in the commanding menu, for injured soldiers regardless if they are incapacitated or not. We can script a Revive or advanced treatment using the HandleHeal event and other scripting. 

 

Just the menu treat option is needed :))

[Gibbo_ 3]

On 2017-08-04 at 1:26 PM, klamacz said:

 

Current plan is to tweak obvious problems till the point that standard skill markers behave similarly with previous implementation. 
As standard skill markers you can consider 0.2, 0.5 and 1.0 skills (so lowest, medium, and highest). We measure reaction times, spotting times, accuracy, decision delays etc and try to get them close to what you seen before with similar values.

That means that conversion table would not be necessary.

 

 

The big change is the curve of skill, we predict that with current code, actual reaction time, or accuracy would be linearly proportional with skill you've set. It means that you won't have to spend time tinkering with your settings for weeks, because you will get more what you actually set.

 

 

Thank you, unbalance will be serious. Also, expect bunch of tweaking commits following that change, and be not too attached to momentary effects you see. Anything might get tweaked as we go through it. 

 

I urge you to consider 100% backwards compability with current values fed into setSkill. Either by a translation table or by some compability flag in the function itself. When you change the underlying algorithm to be more linear and predictable as to how these values are translated into final AI skills, it will nevertheless destroy all tweaking efforts people have made over the years. I have personally spent hundreds of hours on this and would prefer not to have to redo it. 

 

Improvements on AI behavior or tweaking options is appreciated (far more than new DLCs - at least by me), but as im sure you are painfully aware of, it's also very easy to severely break existing missions/mods. 

[Grumpy Old Man 3546]

9 hours ago, lex__1 said:

I have often noticed the following situation when I play on the CTI server. In these missions there are many independent AI. Often AI, in cities, show excellent driving and shooting skills. They drive so fast and dexterously in narrow places (between houses), they are not stopped by obstacles in the form of fences. But at the same time, my AI can stand somewhere near the curb of the road, or at the fork, near the bridge and do not understand - how to overcome them.
But sometimes something happens, independent AIs start to behave very stupid, start to shoot badly, or get a bad reaction. Sometimes there is such a problem, you can go to a group of AI (or any machine with the AI), and they do not react to your presence, you can shoot them one by one (slowly).
Has the intellect disconnected or am I reborn and do not pose a threat to them? In the next city everything is fine, or can happen again in any situation.
I know many places on the mission map where my AI is stuck, with a probability of 100%, but in the AI editor do not get stuck in these places.
I could never reproduce in the editor, any similar problem, to reveal any dependence on the premise of the problem.
This leaves me with an ambiguous view of AI's intelligence problem, or is AI's intelligence code has properties, sometimes stuck in a networked game.

 

 

Is this mission using setFriend by any chance?

 

Cheers

[lex__1 422]

36 minutes ago, Grumpy Old Man said:

Is this mission using setFriend by any chance?

 

Cheers

It seems yes. Three parties hostile - green \blue \red.

[oukej 2911]

18 hours ago, krycek said:

I thought of something else for stuck vehicles. If the AI has problems recognizing it's stuck in a house/tree then why not adding a timer and a check for it's current position until it can initialize backing up?

For example "for 15-20 seconds it seems I keep going forward/accelerating towards my position yet I'm still in the same place,bzzzt,time to back up". I assume they have better detection for other AI vehicles when backing up unlike scenery.

A fail-safe is already present. An issue might be if after backing up the vehicle re-plans via the same route.

[kremator 1065]

I've never seen a failsafe for this stuck vehicle!  There is most definately a need for them to replan as currently they will stay there driving forward for ever.  Please look into it again - as something is severely broken as it stands - which makes vehicles useless!  Just make them use reverse!

Edited August 9, 2017 by R0adki11
removed capitalisation which is not needed, as per the forum rules

[iV - Ghost 50]

Dead Bodies:

Ai should be alerted if seen (new) dead bodies from own side.

They should search the area for the enemy for a while.

[Alwarren 2767]

4 hours ago, oukej said:

A fail-safe is already present. An issue might be if after backing up the vehicle re-plans via the same route.

 

Isn't there something to be done about the re-planning? I've seen this again and again that AI drivers back up and ram the same tree/wall/house again over and over. Something like a temporary memory would be nice to avoid this. 

[dragon01 902]

Actually, it's lack of "memory" that's the problem. AI backs up, then it doesn't remember that there was a problem in a particular spot, so it plots the route using the same principles as it originally did - and arrives at the same conclusion. If it took the "stuck point" into account while re-planning the route, it would have a better chance at getting it right.