AI Discussion (dev branch)

The point relates to the next chunk of work I'd like to invest in the AI. Mr Klamacz is rather tied up with various tasks now but, on the side, he's taking a closer look at the skill configuration and how it really manifests in game.

The goal will be to overhaul the way configuration options are presented to different groups - players and mission makers - giving us a more predictable test-bed and mission makers a more meaningful (and better described) set of configuration options. I'll share an extract from the initial analysis. Please share any general reactions.

Work is a bit delayed on this for a few reasons but the thought is that it will be, in the long run, beneficial to finally lock down these fundamental aspects of AI so that the ongoing tweaks and refinements can be implemented and tested with more confidence. The AI (like some other aspects of the game: video options, etc) is really an incredibly powerful beast, but massively complex to fine-tune. We need to cast a splendid new light on the basic inputs to better unlock the potential of the AI in the wild. :)

Best,

RiE

This is what is really needed. One thing would be clearing up how the skill settings in the game profiles interact with the mission makers skill settings in the editor. Then having more fine tuning ability to control specific parts of the AI's logic, for example I like for the AI to be smart and put fire on the enemy, but I don't like them to be laser accurate because then fire fights are over in 20 seconds.

One thing I would also like to see is the ability to better control formation by shrinking and expanding it.

Then having more fine tuning ability to control specific parts of the AI's logic, for example I like for the AI to be smart and put fire on the enemy, but I don't like them to be laser accurate because then fire fights are over in 20 seconds.

Exactly, I've been cautious to change (lower) any skill settings because I don't wan't "crippled" AI, but I don't like to be killed by an enemy with one or a couple of shots, while I have difficulties to even see the enemy with the same equipment that far.

But I do want him to be able to kill me if I'm able to kill him, if we changed places. Fine tuning stuff.

Sadly.

One thing that bogs me (since i masochistically still play as a AI group leader) is that whenever i command my whole group to move to some spot, while in combat mode, they do move there in perfect formation (wedge) ignoring cover and as walking as slow as possible (not to break that beautiful formation). They skip bounding overwatch alltogether, or at least they fully prioritize hard formation.

Obviously this kind of behaviour is useless.

This was implemented in arrowhead because when they bounded, everybody complained that they "did not move" and "were too slow". So, this works as intended. The other choice, again, was voted out on the basis of the AI being perceived as too immobile in this mode.

yea we basically need sliders for all the parts of the skill array. you can only influence accuracy already but that is limited to your own missions. it would be great to be able to set that stuff quickly when hosting any mission.

I think we need, different AI combat abilities, like a civilians with weapons, should not be so tactical and accurate with weapons, like trained soldiers. There is already some variety between regular soldiers and special forces, but needs a bit more. For example, civs should use "novice ai skill", and special forces "expert"

The goal will be to overhaul the way configuration options are presented to different groups - players and mission makers - giving us a more predictable test-bed and mission makers a more meaningful (and better described) set of configuration options. I'll share an extract from the initial analysis. Please share any general reactions.

Good news :)

One thing would be useful to know, is if discreet AI behaviours are "unlocked" (available to the AI) from a certain threshold upwards, or if they follow a probabilistic likelihood curve of being used by AI. The point is more relevant if skill settings are to be sectioned into levels. (ie. AI never uses smoke in LIMITED, but it does from DEFAULT upwards / likelihood of AI leader initiating an attack (flanking) manouver depending on setting). A succint description shown in user settings should cover these.

As for mission-makers a more detailed breakdown of what purpose each skill setting serve (currently available to an extent), but most importantly more explicitely the correlations between settings (ie. a higher spotdistance combined with lower spottime implying, if any, a qualitative difference in range of AI behaviours). Revealing what, at a more abstract engine level, such settings will mean the AI will do or not do, so that a mission maker can more adequately choose for one setting over another depending on his objectives in a battle.

Edit:

I don't think it is a good idea to absolutely hide the option from the user to independently set friendly AI from enemy side AI. At worse that possibility sould remain available in deeper UI level. The point is that this allows to remove or add player load in working out the encounter when facing AI. This is almost like a difficulty lever on lone-wolfing against tactical play, which is good in accomodating different types of gameplays. Some times i like to "lone-wolf" it out, other times i like to stress the tactical pace and the teamwork.

...unlock the potential of the AI in the wild

Keep it comming.

One thing I would also like to see is the ability to better control formation by shrinking and expanding it.

So much this. Formation or the AI's incessant need to keep it while their lives are in Danger needs serious attention. This alongside with better orientation will go a long ways.

One thing I would also like to see is the ability to better control formation by shrinking and expanding it.

^^+1

One thing i found when messing with the FSMs is that basically AI pathfinding appears to be very determined by a "searchparameter" (ie. [5,10] or [0,0] - meters away on x/y), assuming this is as in "searching for cover" around the formation position around a group leader. What i noticed is that i could statically increase these values and obtain very loose formations while in group movement, which in some instances would prove adequate to circumstances and provide for what is asked.

I would then extend this request to the ability of setting this search parameter dynamically via script aside of allowing a player to do it presumably through some command menu option.

^^+1

One thing i found when messing with the FSMs is that basically AI pathfinding appears to be very determined by a "searchparameter" (ie. [5,10] or [0,0] - meters away on x/y), assuming this is as in "searching for cover" around the formation position around a group leader. What i noticed is that i could statically increase these values and obtain very loose formations while in group movement, which in some instances would prove adequate to circumstances and provide for what is asked.

I would then extend this request to the ability of setting this search parameter dynamically via script aside of allowing a player to do it presumably through some command menu option.

Wow, this looks promising.

I hope one of the sliders will be

Taking Cover

1(Take cover if someone steps on a twig) ----5 (Take cover at the sound of any gunfire/explosions) ---- 10 (march on to the objective while ignoring any gunfire)

Resuming Positions

1(Stay cowering until the next day) ---- 5 (try to zero in on targets and shoot, advance when safe) ---- 10 (Run and gun the whole way until you've slaughtered every thing that moves)

My current area of greatest concern regarding AI skill settings is that there are many different places where the various AI skills are set - within player UI, within the editor, within mission scripting, within addon scripting - and the relationships and interactions between these various settings is largely unclear. If that could be documented and explained it would go a long way to clearing up much of the current confusion.

One thing I would also like to see is the ability to better control formation by shrinking and expanding it.

Made which could help the above: (please vote for it)

0015297: Ability to dynamicaly set searchPath parameter values (via scripting or command menu)

While i can see how this would be useful in achieving the effect of "shrinking and expanding" the formation. It is not exacly the same thing. For that an actual formation "scale factor" would be more specific.

With the ticket i have in mind mostly the situations where the AI is looking too widely for cover while the player (as group leader) asks for the group to retreat or move elsewhere, delaying it's movement to the detriment of it's security as a consequence.

I just played around a little with CAS (Bombing Run) modules in the editor and noticed that the sound of bombs exploding near you on the ground have the same amount of delay as if you were actually sitting in the aircraft. So you hear it about 2 seconds after the actual explosion sometimes there is no sound at all. Explosions also look more like a cloud of smoke actually than a real explosion. Has this issue been addressed already? (and the fact that 9 out of 10 times the aircraft will fail to drop a bomb at all and disappear without a trace)

edit.

this wasn't supposed to be in this thread, sorry

I just played around a little with CAS (Bombing Run) modules in the editor and noticed that the sound of bombs exploding near you on the ground have the same amount of delay as if you were actually sitting in the aircraft. So you hear it about 2 seconds after the actual explosion sometimes there is no sound at all. Explosions also look more like a cloud of smoke actually than a real explosion. Has this issue been addressed already? (and the fact that 9 out of 10 times the aircraft will fail to drop a bomb at all and disappear without a trace)

edit.

this wasn't supposed to be in this thread, sorry

Most bomb explosions are nothing but a cloud of smoke and debris. The actual flash is very very fast. I wish they could look something like this though.

---------- Post added at 21:26 ---------- Previous post was at 21:20 ----------

Made which could help the above: (please vote for it)

0015297: Ability to dynamicaly set searchPath parameter values (via scripting or command menu)

While i can see how this would be useful in achieving the effect of "shrinking and expanding" the formation. It is not exacly the same thing. For that an actual formation "scale factor" would be more specific.

With the ticket i have in mind mostly the situations where the AI is looking too widely for cover while the player (as group leader) asks for the group to retreat or move elsewhere, delaying it's movement to the detriment of it's security as a consequence.

I voted that up. I would also like to see some redundancy in the AI, like rather than being ordered to take cover, they dynamically take cover when stationary within reasonable distance of their formation, say like 10-20 meters. As it is right now, when you order them to take cover, they just kind of hold there forever. Ideally they should do that automatically when in say aware mode or danger mode and they have knowledge of any unknown contact, and still try to maintain formation dynamically. So if the leader moves up say 100m, they will then try to move up into formation while still seeking cover. If they are being shot at then they maintain cover no matter what until the threat is neutralized.

Kind of what the WW AI mod and TPW's EBS mod do in conjunction with one another. I have a feeling that an in engine implementation within the AI's behavior would lead to better performance and better abilities for the AI.

Also the objects they use for cover need to be looked at. I've seen them try to hide behind a tiny light post as cover in the middle of a road. I've also seen them try to hide behind wire fences and such. They also tend to think that they are in cover, even though they are out in the wide open, simply because there is a bush 5 meters from them and it's blocking LOS to a known contact even though 5 other contacts have perfect LOS to them.

I hope one of the sliders will be

Taking Cover

1(Take cover if someone steps on a twig) ----5 (Take cover at the sound of any gunfire/explosions) ---- 10 (march on to the objective while ignoring any gunfire)

Resuming Positions

1(Stay cowering until the next day) ---- 5 (try to zero in on targets and shoot, advance when safe) ---- 10 (Run and gun the whole way until you've slaughtered every thing that moves)

The devs seem to be wanting to move away from player-controlled sliders, so the values a given group takes could be dependent on force size/power. Large groups or groups with elite units would be more aggressive in their response to an ambush (especially the resuming positions part), while a small patrol that just got ambushed would be less so.

Regardless of what options and how many different skill selection criteria is opened up it should simply work something like so:

Mission maker sets a particular value of skill to say 50% for a particular unit. The user has their overall game setting for a particular skill (or perhaps it is just a general overall skill setting of 50%). This equates to the skill setting for that particular unit to 25%.

I am happy to hear that BIS is making steps towards making the ai easier to tweak. However it is not the solution. Basically the problem isn't that the ai is too accurate or not accurate enough, its that their accuracy isn't effected by the events around them. For example ai that has just been hit or is getting shot at should be shooting much less accurately but in higher volume than an ai who is not taking any fire at all. Another example is an ai that has just spotted an enemy should be less accurate than one that has taken a couple shots already and adjusted for his mistakes. Or Ai with magnification should be more accurate but slower to fire. Ai shooting at partial or fully concealed targets should be less accurate. You get the point. BIS has already made it so in CQB surprised enemies will shooter faster but less accurately and this is a great step in the right direction - it shows in gameplay. I believe that ai initial shots are less accurate at first as well- but this needs to be extended to when you leave their los and pop up in a new location to account for the fact that they have to adjust their aim once more.

This would truly make interesting firefights. If the ai's accuracy always remains static decreasing their skills through sliders will only make the game easier - it will not make the game realistic or have any more need for tactics.

If you don't believe consider this:

• Situation 1: ai gets you in an ambush from 100m - Realisitcally your squad should be for all intents and purposes decmiated.
  
• Situation 2: you and the ai meet in a suprise contact at 100m - Realistically your squad and the ais would probably survive long enough to get to cover.
  

with the current system only one of these can be possible - if you decrease the skill, suprise contact will be realistic while ambush won't be - too forgiving. If you increase the skill the ambush will be realistic but the surprise contact will not - would be far too quick. I hope you get the point.

So While I am very happy to see bohemia making the ai accuracy more configurable I really hope it is just the first of many steps to fixing the "accuracy problem".

I hope one of the sliders will be

Taking Cover

1(Take cover if someone steps on a twig) ----5 (Take cover at the sound of any gunfire/explosions) ---- 10 (march on to the objective while ignoring any gunfire)

Resuming Positions

1(Stay cowering until the next day) ---- 5 (try to zero in on targets and shoot, advance when safe) ---- 10 (Run and gun the whole way until you've slaughtered every thing that moves)

Well of course this would be good and it would be even better if mission designers could set it but first it would require Ai to actually know how to take and use cover in the first place and recognize when it is safe or unsafe. Right now Ai cover taking is spotty at best and they certainly don't know how to duck in and out of it like a human would to limit exposure.

But yes, it would be cool if you could set how willing ai are to take cover and how far out of formation they go to do so and how long they willl do it for.

Mission maker sets a particular value of skill to say 50% for a particular unit. The user has their overall game setting for a particular skill (or perhaps it is just a general overall skill setting of 50%). This equates to the skill setting for that particular unit to 25%.

Yes this is how it should be done.

But this is one thing and what they're doing is other; cleaning the mess in configurations (skill, precision, skill in the editor +++ User CPU) both for the user and for them to "narrow" the AI behavior. Once they know what the AI is thinking (or proned to), they can tweak the FSMs.

However it is not the solution.

Of course not... it is not meant to be a solution to the problem you're focusing on. If you notice Bohemia is trying to takle another problem here: AI configuration settings aren't very clear to players / Players can unwittingly 'break' the game by setting values that are too low / Mission designers can't easily modify useful parameters of AI. If you ask me, i also find a solution to correct/acceptable AI behaviour more necessary, yet addressing the this should allow for higher predictability and stabilization in resulting behaviour, which could go a long way in addressing the broader AI problem.

Edit: Just think of the feedback on AI issues, underlying difficulty, and how it reaches Bohemia... it is all over the place. Definite levels of difficulty instead of an array of AI values, should keep it to a much controled/useful to whoever is programming and correcting AI issues.

Mission maker sets a particular value of skill to say 50% for a particular unit. The user has their overall game setting for a particular skill (or perhaps it is just a general overall skill setting of 50%). This equates to the skill setting for that particular unit to 25%.

Equally agreed, user should be exposed to a multiplier of whatever AI difficulty/balance is set upon mission (instead of overwriting it). Be this user setting linear difficulty values or accessible via discreet levels.

Edit2: I see the above creating a basis for a guideline for mission-makers to follow. The example mentions a general skill of 50%, but maybe a better recommendation was for missions to be created with 100% skill in use, so that the full capabilities of AI are tested under that circumstances by the mission maker, then the user would apply its intended setting/multiplier. (Note that, as i see it, it is not very likely that a user general multiplier would degrade the AI necessarily in linear terms on all it's different settings)

But this is one thing and what they're doing is other; cleaning the mess in configurations (skill, precision, skill in the editor +++ User CPU) both for the user and for them to "narrow" the AI behavior. Once they know what the AI is thinking (or proned to), they can tweak the FSMs.

If you ask me, i also find a solution to correct/acceptable AI behaviour more necessary, yet addressing the this should allow for higher predictability and stabilization in resulting behaviour, which could go a long way in addressing the broader AI problem.

Edit: Just think of the feedback on AI issues, underlying difficulty, and how it reaches Bohemia... it is all over the place. Definite levels of difficulty instead of an array of AI values, should keep it to a much controled/useful to whoever is programming and correcting AI issues.

Oh yes that's why I say

So While I am very happy to see bohemia making the ai accuracy more configurable I really hope it is just the first of many steps to fixing the "accuracy problem".

Its just that I have heard many times people say stuff like "all we need is a slider to set the ai accuracy to our liking." I just hope BI and people realize that this is not all we need. That's all. Its a good point you make about feed back being all over the place. Hopefully this will give BIS a better idea of "whats up". I also find that this script is good for tracking how accurate ai is while observing using the splendid camera in slow motion.

Regarding the discussion about how granular AI options need to be I would like to drop my humble opinion:

1. I think all tactical capabilities of the AI should always be at the best level the engine can provide.

This covers things like caring for their own lives / suppresssion, choosing the right weapon for a situation, flanking, handling of formations, searching in the right places, taking cover, driving without crashing, rearming, applying first aid to oneself and ones comrades, ...

This aspect is, which can make AI appear human-like, if it is implemented reasonably, but which also has the potential to totally ruin gaming experience and break immersion if done poorly.

So I would always use the best the engine can give us, and never dumb down AI by disabling or crippling capabilities.

2. It would be completely sufficient to tune AI shooting accuracy and reaction speed for adjusting game difficulty to ones personal liking.

I know about discussions like specific needs of mission designers for scenarios like "Super Human Special Forces vs. a bunch of retarded civilians with fire arms", but I think these can really be dealt with by 2.

IRL even completely untrained and stupid civilians would behave more cleverly than the current level of gaming AI. And this statement is not meant to blame BI. I think the ArmA AI is at least one of the best we can currently see on the market. AI is simply the biggest challenge in designing computer games.

For resembling different training levels and shooting experience it should really be sufficient to simply tune down shooting accuracy (and maybe reaction and detection speed).

So, what is your opinion?

I did some test yesterday with the actual Dev Built. I had to reverse my own "impressions" after the test results.

http://forums.bistudio.com/showthread.php?166487-ARMA3-Coop-Tactics-how-to-fight-the-AI&p=2530218#post2530218

Regarding the discussion about how granular AI options need to be I would like to drop my humble opinion:

1. I think all tactical capabilities of the AI should always be at the best level the engine can provide.

This covers things like caring for their own lives / suppresssion, choosing the right weapon for a situation, flanking, handling of formations, searching in the right places, taking cover, driving without crashing, rearming, applying first aid to oneself and ones comrades, ...

This aspect is, which can make AI appear human-like, if it is implemented reasonably, but which also has the potential to totally ruin gaming experience and break immersion if done poorly.

So I would always use the best the engine can give us, and never dumb down AI by disabling or crippling capabilities.

2. It would be completely sufficient to tune AI shooting accuracy and reaction speed for adjusting game difficulty to ones personal liking.

I know about discussions like specific needs of mission designers for scenarios like "Super Human Special Forces vs. a bunch of retarded civilians with fire arms", but I think these can really be dealt with by 2.

IRL even completely untrained and stupid civilians would behave more cleverly than the current level of gaming AI. And this statement is not meant to blame BI. I think the ArmA AI is at least one of the best we can currently see on the market. AI is simply the biggest challenge in designing computer games.

For resembling different training levels and shooting experience it should really be sufficient to simply tune down shooting accuracy (and maybe reaction and detection speed).

So, what is your opinion?

^this

Found this reaction to the dev branch changelog item about AI leading of moving targets from turrets getting improved/buffed from off-forums:

That's OK, facing a T100 wasn't horrifying enough in the previous builds, I mean, there are parts of my chair that aren't covered in feces yet.