The Schizophrenia of the Tac AI

[mjkerner]

The Smart Tac AI

Pixeltruppen: Hey Boss, there’s some enemy guys over there by the trees! Imma fire at them.

Boss: Where, I don’t even see nuttin?

Pixeltruppen: Pew-pew-pew-pew-pew-pew!

Boss: What the goddam hell are you firing at?

Pixelpanzer: Hey Boss, I see some enemy guys over there. Imma turning my turret to fire at em!

Boss: Where, I don’t see nuttin’?

Pixelpanzer: Bam!

Boss: Goddam it to he—

Pixelpanzer: OUCH!  Imma pop smoke an’ back in’ da flock outta here! LIKE NOW!!!!

Boss: What the hell? Where? What?

Pixeltruppen and Pixelpanzers: <sotto voce> God, he ‘s stoopid and blind!

 

The Dumb Tac AI

Boss: Okay, you guys go over there, then slip through that break in the bocage and quicktime it to that building across the field.

Pixeltruppen: Shooer ting, Boss!

Pixeltruppen: Okay Boss, the eight of us guys are through the gap and headed for th——

Enemy Bastiges: Brrrrrrrrrrrrrrrp! Brrrrrrrrrrrrrp! Brrrrrrrrrrrrrp!

Pixeltruppen: Boss, it would likely appear to an observer that we are taking fire. Are we? No matter, we see the building from here, so we’ll just sorta keep moseying on over there...

Boss: NO, NO, get down! Hit the dirt! LIKE NOW!

Enemy Bastiges: Brrrrrrrrrrrrp, Brrrrrrrrrrp, Brrrrrrrrrrrp!  <Hey Boss Man, we can and will keep this up all day ! You and your Pixeltruppen are stooopid!>

Pixeltruppen: Hey Boss, the five of us think we still can make, and Jones is going to stop—-standing up—-and pew pew back!

Enemy Bastiges: Brrrrrp! Brrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrp! Brrrrrrrp!

Boss: Hit the Phuuuuking dirt NOW, you IDIOTS!

Enemy Bastiges: <Ha, ha! He said “Phuuuuking!> Brrrrrrrrrrrrrrrrp!

Pixeltruppen: No sweat Boss, we don’t mind you swearing! Smitty an’ I can probably still make it to th———-aaarrrgh!

Enemy Bastiges: Get some! Get some!  

[The above was an actual exchange during a CMx2 firefight. Names have been withheld to protect the guilty Boss Man.]
 

I don’t so much mind the enemy taunting the Boss, and my truppen dissing him, but why is the Tac AI so schizophrenic? I know its been complained about ad nauseum, but I don’t recall much discussion, if any, about this apparent disparity within the TacAI system causing it. Under it, static (and to a degree, moving) troops spot, react to threats, and fire intelligently for the most part. Programming rules are obviously in place to cover these situations. But when under Movement Orders, there seems to be a different set of rules. Or what, exactly?  Moving, they still will spot threats, react, and even countermand our orders—(channeling Newt here) mostly—but rarely if ever hit the dirt when under fire, until the end of the order. There must be some sort of Orders override regarding static troops—cowering, running away, refusing to initiate your move order, etc., when the **** hits the fan. But nothing like that seems to be in place when moving under heavy infantry fire, or worse, an artillery barrage. 
 

I know the arguments regarding the difficulty in making programming decisions about this, like at what point /how many casualties to take before the TacAI says “hit the dirt” is initiated, etc. Players’ reactions to not being able to control their troops like we would like are often mentioned as one main reason to not implement a “go to ground” command.  But hey, that doesn’t stop the TacAI from refusing to move troops, or cowering, and other behavior that we lose control of when preservation is at stake.

Anyone have insight that can be shared?

[Erwin]

Programming AI is the most complex task in making a game - hence the preponderance of online multiplayer games that require crappy or no AI.  Yes, all of the above you mentioned is frustrating.  With experience, and the understanding that one is playing vs the game system much of the time, one simply learns to work around it.

[chuckdyke]

3 hours ago, mjkerner said:

Anyone have insight that can be shared?

Risky moves? Pause for up to 45sec before committing. Start on hunt 1 or 2 squares before dashing across. Don't ignore the stress levels, it starts with green otherwise it is blanc don't ignore the green once it is yellow or red you lose control of the unit. OODA loop find it a big help. Observe, Orient, Decide and Act. Stress level red hand it over to the AI by pressing the emergency button. Squad often recover from stress by splitting or split squads by rejoining. 

[Erwin]

20 minutes ago, chuckdyke said:

Risky moves? Pause for up to 45sec before committing.

Yes, this is good advice.  It's a shame that we cannot choose longer or shorter time intervals for turns as this method can use up a lot of time if done repeatedly.  However, am sure that the AI couldn't cope with that.  However, in H2H play, it would be great to be able to agree with one's oppo on (say) 30 sec or 90 sec moves.

[mjkerner]

Thanks fellas, but the insight I meant wasn't how to handle it in the game, but why is it that it can work well in one instance and not the other. I think it is more of a question for BFC/Steve/Elvis, but thought maybe some of the long time beta testers can weigh in. Sorry I wasn't very clear.

[Freyberg]

'Max Assault' will allow the AI to stop and respond to threats, even to the point of not reaching its target zone.

Seems to work OK...

[domfluff]

If I'm following the actual question, then that's essentially "Why do troops react differently when moving and they come under fire".

The design intent is really part of the underlying logic. CM offers you a number of orders, and they all have some underlying logic to it, which might not be entirely intuitive.

As a simple example - "Face" isn't just "turn in that direction", it's more like "be prepared to fight in that direction", meaning that the Face command cancels cover arcs, turns the troops to face, and they'll shuffle into covered positions where they can, taking up positions against a wall, behind trees, etc.

The movement commands then are a trade-off between speed and caution. "Fast" isn't just "move over there quickly", it's "Your priority is to move to that location as fast as possible, to the exclusion of all else" -  Fast moving troops should rarely stop to fire, and instead they're focused on sprinting. "Fast" then is ideal for crossing open streets, or getting away from incoming mortar rounds - anything where being precisely here is a bad idea.

The movement orders then all represent this trade-off, speed for security. Hunt-Move-Quick-Fast represent the spectrum between maximum caution (and stealth) and maximum speed. "Slow" is really "Crawl". It is the slowest and most cautious, but the crawling is the thing, since it maximises stealth and minimises exposure when cresting a hill or whatever. This control is really important - there are situations where you really need one or the other, and having these options is important for simulating real behaviour.

Now, wherever you draw the line it's going to be wrong in some way - there's arguments to be made about whether the AI is operating reasonably for the incoming stimulus. What makes that more complex is that what you see and what the squad are aware of are not necessarily the same thing.

[Ultradave]

Not really helpful but the first conversation sounds a lot like reality

 

[mjkerner]

domfluff, I understand all that, as with Chuck’s and Erwin’s responses, cuz I have been playing and delving into the innards of the game pretty much daily for 14 years.  I guess I am asking the question wrong, or maybe asking the wrong question. I think the “answer” is “forget about it, it’s all under the hood”, lol. 

What I’m trying to get at is why does the TacAI override your given orders sometimes, and at others, slavishly try to follow them. Specifically, in the case of enemy AI continually running through or into a kill zone without stopping. (And to a lesser degree, why do my men sometimes not hit the dirt when fired on, even in Quick or Hunt mode.) One answer would be because the game checks soft factors by individual unit (as chosen by you or the scenario author for the enemy at the moment of calculation i.e., is this only a fire team or a whole squad?).  Therefore, the enemy squad running into a kill zone will hit the ground after X volume of fire/casualties taken, but the next individual unit has to take the same check, so naturally, they will continue moving into the zone until their factors are checked against the incoming fire/casualties taken. All understandable to me so far, cuz the game doesn’t tell the unit’s commander (in computer language) “Hey, your guys are getting slaughtered...stop sending them through that kill zone!”

Only it does, to some degree.

When you split squads, or move units away from each other or its leader, and the parent unit (squad if it’s a fire team, platoon if a squad) takes casualties or some similar calamity happens, it does affect the survival behavior of the split-off unit. They will cower or run away...whatever. That behavior is greatly affected by soft factors, to be sure, as well as the the specific orders given, as Freyberg points out.  But in effect, the TacAI is deliberately cancelling the human’s orders, and essentially issuing its own for that immediate case. I guess I was just trying to figure out why it doesn’t go one step further and issue orders for the parent unit to hit the dirt and/or take a different path?

Just keying all this out, I guess I found my answer. Having the TacAI read the situation and issue new orders to move here, or hit the dirt there, is already covered by the interplay of orders, soft factors, etc. I certainly don’t want the AI to override my orders (or the scenario  author’s) each and every time heavy fire/heavy casualties is taken. Regardless, my hat is and has always been tipped to Charles Brain -in-a-jar, Phil Culliton(sp?), and whoever else figured out the programming for CM. Man, my mind aches just trying to ask this question!

It is a wonderment.

 

[Vergeltungswaffe]

Programming tactical AI responses in a game has to be the hardest undertaking in computer gaming. Strategic, not so much.

But, as has been proffered, if anyone ever programs an AI that can play a game like a human, we're all toast.

[markshot]

I just finished some MIUS.  CM handles psych states better.  Which is not to say they are ignored in MIUS.  But playing CM feels like an emotional roller coaster where as MIUS feels like a simulation of mental states.  I don't think you can get both a human like simulation of emotions without getting some of what appears senseless activity.  (It is only senseless when seen outside by a player who knows its a game.)

The thing that used to bug me was a few guys always finding stupid places to stand or go.  Like your buddy just got hit standing on the sewer cap.  So, where should you stand?  Of course, on the sewer cap.  I think 4.02 is doing better with that.  I remember posting why do 2 guys always have a death wish when fire is coming across a wall, and they always want to check out the other side.

[nik mond]

7 hours ago, Ultradave said:

Not really helpful but the first conversation sounds a lot like reality

 

That first conversation sounds like...Orcs

[Vergeltungswaffe]

16 minutes ago, nik mond said:

That first conversation sounds like...Orcs

I read it as 40's vernacular, but now that you've said that. All I hear is Orcs.

It just needs a WAAAAAAGGGHHH! when they go through the bocage gap.

[domfluff]

22 hours ago, mjkerner said:

domfluff, I understand all that, as with Chuck’s and Erwin’s responses, cuz I have been playing and delving into the innards of the game pretty much daily for 14 years.  I guess I am asking the question wrong, or maybe asking the wrong question. I think the “answer” is “forget about it, it’s all under the hood”, lol. 

What I’m trying to get at is why does the TacAI override your given orders sometimes, and at others, slavishly try to follow them. Specifically, in the case of enemy AI continually running through or into a kill zone without stopping. (And to a lesser degree, why do my men sometimes not hit the dirt when fired on, even in Quick or Hunt mode.) One answer would be because the game checks soft factors by individual unit (as chosen by you or the scenario author for the enemy at the moment of calculation i.e., is this only a fire team or a whole squad?).  Therefore, the enemy squad running into a kill zone will hit the ground after X volume of fire/casualties taken, but the next individual unit has to take the same check, so naturally, they will continue moving into the zone until their factors are checked against the incoming fire/casualties taken. All understandable to me so far, cuz the game doesn’t tell the unit’s commander (in computer language) “Hey, your guys are getting slaughtered...stop sending them through that kill zone!”

Only it does, to some degree.

When you split squads, or move units away from each other or its leader, and the parent unit (squad if it’s a fire team, platoon if a squad) takes casualties or some similar calamity happens, it does affect the survival behavior of the split-off unit. They will cower or run away...whatever. That behavior is greatly affected by soft factors, to be sure, as well as the the specific orders given, as Freyberg points out.  But in effect, the TacAI is deliberately cancelling the human’s orders, and essentially issuing its own for that immediate case. I guess I was just trying to figure out why it doesn’t go one step further and issue orders for the parent unit to hit the dirt and/or take a different path?

Just keying all this out, I guess I found my answer. Having the TacAI read the situation and issue new orders to move here, or hit the dirt there, is already covered by the interplay of orders, soft factors, etc. I certainly don’t want the AI to override my orders (or the scenario  author’s) each and every time heavy fire/heavy casualties is taken. Regardless, my hat is and has always been tipped to Charles Brain -in-a-jar, Phil Culliton(sp?), and whoever else figured out the programming for CM. Man, my mind aches just trying to ask this question!

It is a wonderment.

 

Well, that's part and parcel of a simulationist design.

There have been very few attempts at true determinism in wargame design, and the ones that have (the games of Bowen Simmons are perhaps the most notable in recent years), have replaced the friction of die rolling with the friction of hidden information, which can be mathematically identical.

Starcraft has units which will follow the precise inputs reliably, but Starcraft is gamist in intent. The point of a sim is frequently *not* to have these kind of slavishly repeatable results, because that doesn't reflect the reality of the player role simulated. In the case of CM, broadly everything platoon leader and up. A platoon leader will be doing a pretty poor job if they were attempting to micro-manage the location of each individual and weapon system in their platoon.

So, if the question is "why does this work this way mechanically", it's as I said - you're juggling tendencies between security and speed. Not firm numbers, but probabilities.

If it's "why does it modelled in this manner?" then that's the kind of abstractive line that wargames have taken since Kriegsspiel - you're relying on the soldiers to do their thing, anyone in the role of a platoon leader or higher would do.

The latter is a significantly more interesting question, since there is far from one possible answer, but philosophically the role of variance in wargaming dates back to at least Clausewitz. 

[Jabble]

On 4/27/2021 at 11:51 PM, mjkerner said:

Anyone have insight that can be shared?

More appropriate reaction to events would be desirable.

On 4/28/2021 at 5:24 PM, Vergeltungswaffe said:

Programming tactical AI responses in a game has to be the hardest undertaking in computer gaming. Strategic, not so much.

But, as has been proffered, if anyone ever programs an AI that can play a game like a human, we're all toast.

I reckon the Tac AI here could be improved by a configurable "react" posture that would be automatic during WeGo resolution, e.g. hit the dirt, fight back, and/or retreat.  That way troops could be instructed to behave more realistically when reacting to events during that golden minute.

There's an inherent difficulty in the WeGo approach where the main intelligence is the player who only sees what he sees during the command phase, and the pixeltruppen who tend to behave like unthinking robots, slavishly following those orders given less than a minute ago. That's certainly a challenge for AI to mitigate but I do think a few basic reactions would make an enormous difference there.  It could work just the same way for the AI side so no imbalance there.

[herr_oberst]

6 hours ago, Jabble said:

  It could work just the same way for the AI side so no imbalance there.

But while you as a player could tailor a certain squad's, or a set of squads, 'react' behavior according to your instincts or plan, choosing to be cautious or charging forward losses irrelevant, how is the tac AI going to make the same decision on setting the 'react' posture for its planned moves?

Some type of algorithm, which I assume is already in place when the tac AI chooses between the slow--fast movement spectrum for its pixeltruppen.

One of the things I've always liked about the CM titles is the lack of precise control once you get far enough down the command chain. That 'randomness' keeps the same scenario from playing the same every time.

[Jabble]

9 hours ago, herr_oberst said:

But while you as a player could tailor a certain squad's, or a set of squads, 'react' behavior according to your instincts or plan, choosing to be cautious or charging forward losses irrelevant, how is the tac AI going to make the same decision on setting the 'react' posture for its planned moves?

Some type of algorithm, which I assume is already in place when the tac AI chooses between the slow--fast movement spectrum for its pixeltruppen.

One of the things I've always liked about the CM titles is the lack of precise control once you get far enough down the command chain. That 'randomness' keeps the same scenario from playing the same every time.

I'd suggest a set of defaults that would apply to both sides, based on relative threat as well as unit abilities. On the AI side this could be varied by a random 'commander demeanour' factor, from cautious to gung ho.  The AI may also be influenced by any overall stance or goal prioritisation at that time, e.g. prepared to take higher risks to push through a high-priority objective.  There could also be a surprise factor that should give some sort of bonus to the unit less affected, e.g. on a 'hunt' move.

I envisage that a 'react' would be pretty much the same set of behaviours you'd expect the AI to carry out during the command phase. I'm not necessarily talking about any new functionality - I'm suggesting that it would be a trigger to standard game functions, and would relate to dealing with a threat. For example, a scout team encountering a tank would retreat immediately rather than fight, whereas a deployed ATGM team would fire a shot and then relocate ASAP.  MBTs would go toe-to-toe as usual.  In some ways this would be the same as if the encounter happened during the last second of the WeGo minute - the AI would then have to make similar decisions based on this new information.

Perhaps some new features could apply; the human side could allow setting a 'retreat to' location (based on current waypoint) while the AI would have to determine where's the nearest cover at the moment of encounter.  That way an ATGM team could sit tight until a target appears, take the shot, and then scarper to that predetermined location.

I put a post in to the suggestions/requests thread along these lines.

[Norman D Student]

A good discussion, with some interesting ideas.

But I wonder if a basic part of the problem here is the name, 'Artificial Intelligence'. Is that what it really is? Well, yes and no. 

What is NOT being modeled by the so-called AI is some Spock-like beings who manage to be coldly rational at all times, along the lines of the 'idealized rational actors' of so much economic theory.

What IS being modeled is human beings, soft squishy things that go into combat with (a) training at doing certain things preferably without thinking too hard about them, plus (b) variable, hard-to-model individual and group dynamics that can cause them to flip out completely under the stresses of combat. As you go up the ranks, you add more and more of what is more traditionally considered 'intelligence' in the 'AI' sense: (c) observe, assess, plan, execute, etc.

So I like to think of the AI as in fact 'artificial training + [psychological] vulnerability + intelligence' -- that is, as ATVI. 

It seems to me that the I part works pretty well. (I'm an experienced computer programmer, so I'm somewhat aware of what goes into programming AI, though it is not my specialty.) Though I do find that the line-of-sight modeling is a bit dodgy, and that plays a significant role in the I.

Most of the complaints in this thread are about the V, where there are really two distinct problems: When does the V take over? That is, when do the pixeltruppen totally lose it? Second, what do they do when they lose it?

Let's take the second of these first. It is not a simple thing to use a deterministic system (a computer program) to model the behavior of someone who has gone nuts. That said, my limited reading about and zero experience of combat stress does not make me think that CM has got this area quite right yet. For example, I've repeatedly seen pixeltruppen freak out when behind bocage, and then run through the nearest opening to the other side of the bocage -- which is where the fire is coming from that has prompted their freakout. My reading suggests that this is not the most common expression of combat breakdowns. I'm more inclined to think that going into a very long-lasting, fetal-position suppressed state should be pretty common (one reads lots of accounts of that happening on battlefields). From what I see, it isn't common in CM. And that should be followed closely by running AFAP *away* from the enemy. Which does happen a fair bit in CM, so better marks for that. Freezing in a standing position should happen more (I don't think I've seen it happen at all). But running blindly *toward* the danger should, IMO, be much rarer than it is in CM. So I would judge this area as one where there's room for improvement.

Then there is the question of how readily do the pixeltruppen lose their minds. I think the US, Cdn, Polish and Brit forces in CMBN are too psychologically fragile. Haven't commanded the Germans, so I can't say. But read, for example, James Holland's _Normandy '44_, and one gets the impression that the real troops endured a lot more torture before becoming shell-shocked than do the CMBN pixeltruppen.

One specific complaint I would have in this area is that it would be more realistic for combat fatigue to take out individual pixeltruppen (like casualties) than for it take out teams, which seems to be what usually happens.

So: Another area for improvement, I think.

Which brings us to the T part of ATVI, training. This, it seems to me, is an area where CM gets mixed marks. What I've noticed is things like this: Germans are super quick to use grenades in appropriate situations, while Allied pixeltruppen ... aren't. Indeed, it just never seems to cross their minds! I rarely become so unhinged as to talk to my pixeltruppen out loud, but 99% of those times, what I'm saying is, "Throw. Your. Grenades. Pretty please?" and other things to that effect.

Apologies for such a long ramble. But I'd be curious if any of these thoughts ring true to anyone else.

Edited May 1, 2021 by Norman D Student

[Norman D Student]

A quick afterthought: 

You can tell that I'm not very good at CM yet by the fact that I've been doing such a thorough study of pixeltruppen combat fatigue. 

[Vacillator]

20 minutes ago, Norman D Student said:

Apologies for such a long ramble

I like your ramble and I'm happy to read more.  Good post.

[MikeyD]

Players tend to max out their troop stats when purchasing but that has a downside. A 'Fanatic' pixeltruppen  is very much less likely to go to ground when fired on. He'd rather die on his feel then live on his knees, so to speak. Sometimes its desirable for your troops to high-tail it out the back door rather than die in place. Also, when you give them a 'Run' command they're liable to ignore everything else and just run. With a 'Quick' command they're more likely to prioitize staying alive or shooting back at the enemy.

[MOS:96B2P]

34 minutes ago, Norman D Student said:

A quick afterthought: 

You can tell that I'm not very good at CM yet by the fact that I've been doing such a thorough study of pixeltruppen combat fatigue. 

You may find the below topic interesting. 

 

 

[Ultradave]

2 hours ago, Norman D Student said:

I've repeatedly seen pixeltruppen freak out when behind bocage, and then run through the nearest opening to the other side of the bocage -- which is where the fire is coming from that has prompted their freakout. My reading suggests that this is not the most common expression of combat breakdowns. I'm more inclined to think that going into a very long-lasting, fetal-position suppressed state should be pretty common (one reads lots of accounts of that happening on battlefields). From what I see, it isn't common in CM. And that should be followed closely by running AFAP *away* from the enemy. Which does happen a fair bit in CM, so better marks for that. Freezing in a standing position should happen more (I don't think I've seen it happen at all). But running blindly *toward* the danger should, IMO, be much rarer than it is in CM. So I would judge this area as one where there's room for improvement.

 

It's already been improved. That behavior has been programmed out. RT is the last to receive this update. And while it didn't happen as much in RT, during playtesting I was facing some Russian troops in a  farm complex who suddenly rushed toward me. That was before the same fixes that the rest of the games already had were applied as part of the update containing RT. BN was by far the worst affected of this, and I think the consensus is that troops behavior now is much improved and more what you'd expect.

If you ARE still seeing this somewhere, please try to duplicate and provide a save. It shouldn't be happening any more so if it is, it should be investigated.

Dave

Edited May 1, 2021 by Ultradave

[General Jack Ripper]

On 4/27/2021 at 6:51 PM, mjkerner said:

Snip

Context needed.

What orders did you use?
How far away was the shed?
Did you recon by fire the field first?
Did you pop any smoke?
Did you split your squad into teams?
Did you probe with one team first, then order the rest to follow?
What is the location of the enemy?
What direction is the squad taking fire from?
Were they detected before you started your movement?
What is the experience/morale of your men?
What time of day is it?
Did they eat lunch before setting off?
Did you have a few beers before playing?

The answers to these and many other existential questions could help us drill down to a plausible answer, as a result of my own experience playing this game.
Here's hypothetical example:

If the squad was just given a Quick command through the hedgerow to the shed, it is likely there was no enemy detected before they set off. The squad would have gotten strung out passing through the hedgerow gap, and when they came under fire some men were outside the hedge, and some were still inside it.
Upon being taken under fire initially the squad might not have received enough suppression to stop moving, and so the guys outside the hedge kept moving toward the shed. As the enemy fire builds up quickly they realize they're in danger and decide to stop. One or more men might now decide to return fire.
Some men might panic at this point realizing they're in the open, outgunned, and in danger. Some might run back to the hedge to join the rest of their buddies, the rest might decide it's faster and easier to move to the shed instead. Some might just decide to go prone, curl into a ball, and pray.
The result of this is that the squad gets divided and some guys get shot for their trouble, and it's all because the Sergeant is an idiot and made no preparations for moving through potentially hostile open space by doing some recon, or popping some smoke, or splitting the squad into teams and sending one team forward while the rest stay back and observe or react to enemy contact.

The TACAI can only try to have your men react AFTER contact is made, everything that happens before that point is entirely up to you, the player.
If you properly prepare the ground you fight on, you won't need an AI helper to fix your mistakes, because you won't make them in the first place.

I hope this helps. Thanks for your time.

[chuckdyke]

In the game I took fire 2 men were wounded. Did all the right things, when I seized the building, the fire came from didn't find any of the enemy. How long does the game display fallen enemy? It is helpful to get intel.