Weak enemy tank AI

[IMP-22]

I don't know if this has been mentioned before, but I've noticed that the enemy tank AI is pretty idiotic. I recently set up a game of 3000 points, giving myself only a crack platoon of Tiger tanks. I set them up behind a couple hills and waited about 10 minutes, then I brought them out and started blasting at the T-34s at long range. The T-34 crews basically just spun around in circles, exposing their vulnerable side and rear armor, while firing harmlessly at my Tigers. They made no effort to actually use the T-34s superior speed and attempt to rush my Tigers. They just clustered up in a big group and spun around in circles, making themselves cannon fodder for my Tigers.

It seems that whenever an enemy tank encounters one of your tanks, and it has superior armor or the gun on the enemy tank is too weak, the enemy tank will slam it in reverse and spin around, exposing its flanks. Once it gets out of view of your tank, it will 'hunt' forward again, then reverse again once it sees your tank. Or get blown to bits, whichever comes first. Enemy tanks never 'fast' their way forward, trying to get in a killing shot or surround your tank. Your own tanks will often reverse as well, if they encounter a tank with superior armor that they know they can't penetrate, but you can tell it to race forward and try to close the distance with the enemy tank. The problem is, the enemy tank AI doesn't seem to ever want to do this, and just keeps reversing.

This problem wasn't as apparent in CMBO, apparently, because the two sides were fairly balanced in most encounters. Having either superior guns or weaker armor, so they didn't spin around in circles like a lot of tanks in CMBB do. For example: A Sherman and a Panzer IV in CMBO were pretty equal in long range encounters, so they would just sit there and blast eachother to bits. But if a T-34 is pitted against a long barreled Panzer IV in CMBB at long range, the T-34 will go into reverse and get itself blown to pieces, instead of charging forward as it should.

So instead of a lone Tiger tank couragesouly beating off a wave of T-34s as they charge up the hill, the Tiger just sits there while the T-34s run into eachother and slowly die. Can't this problem be fixed in a patch or something?

[Sublime]

You're asking too much. The tank AI isnt idiotic all the time, only in specific situations, and I think it'd be impossible to code the AI to do what you want. (at least given what we have nowadays) Even if they could code it to what you want, people would be bitching then about 'The AI always just does tank rushes'

You cant keep everyone happy. There are unfortunate side affects to the AI and how the game is, but if you cant handle it play human opponents. Otherwise, I guess dont play any more scenarios where you'll have a few Tigers facing a horde of T-34s at long range. If they patch anything I think they just need to patch the tank gun firing problem.(with crests and whatnot)

By the way, are you sure about the T-34 vs the Pz IV? My T-34s never seem to back off against Pz IVs...

Also in CMBO the sides were not really more balanced. Certain periods on the Eastern Front saw big disparities in equipment, true, but I'd take T-34s and IS-2s in CMBB over Shermans and Churchills in CMBO...

[BlackVoid]

Most of the time the AI tanks are easy prey for a human player.

It is true that it would be very hard to program a competent tank AI.

What you can do is:

1. Play infantry only games against the computer.

2. Play tank battles against a human opponent.

3. Play only scenarios agains the computer.

[JasonC]

No, IMP-22 is not asking too much. Tank cower behavior is extreme in CMBB, and it is broken.

The main problem is that tank cower behavior does not take into account the facing of the superior "scary" vehicle. This needs to be changed.

At present, it hurts not only the clueless AI but even the best possible human directed play. The Tac AI fumbles the "hand off" for the last 5 seconds.

If you expertly maneuver a T-34/76 with tungsten into a rear shot on a Tiger I at 100m range, with the Tiger facing completely the wrong way, the T-34 will often refuse to take the shot, and instead reverse the instant LOS is established. This is as realistic as a bucket of warm spit.

The present cower behavior should be restricted to vehicles facing a superior AFV that is facing within 30 degrees of them, *and* targeting them. This is not what presently happens. The change would allow proper team tactics against thick AFVs, flanking moves, etc.

The problem is made worse by the fact that cowering AFVs spend time turning, which is slow, and by the fact that vehicles moving in reverse do not fire.

In addition, I see cower regularly when the enemy AFV can penetrate the cowering tank, when the cowerer can also kill the enemy and has the drop on it. I've had IS-2s with properly aligned turrets on vehicle covered arcs repeatedly cower from Pz IVs facing another direction.

This sort of thing should be restricted to cases where there isn't a ready round - meaning, the tank is presently on "reload" delay.

Concretely, if a tank has LOS and the gun or turret is closer to aligned for a shot than the enemy's is, it should first take its shot, and cower only afterwards (if penetrable, and the enemy is targeting, and the enemy is facing within 30 degrees).

If this occasionally leads to a delay and a resulting loss for a slower turreted or slower turning vehicle, so be it. It is better than case after case of complete refusal to shoot, reversing on contact.

This has nothing to do with vehicle morale states, incidentally. It happens with veteran AFVs in command by morale bonus leaders, before a shot has been fired.

The Tac AI is simply programmed to reverse (and pop smoke if available) instead of shooting, when a superior AFV is detected, regardless of facing or whether a round is ready in the tube. And that is the wrong SOP 9 times out of 10.

I won't play scenarios with heavy tanks in them at present, because this issue is so bad that tactics in these situations have nothing to do with reality.

This is one case where it is really a matter of "BTS please fix or do somefink".

[ March 30, 2003, 12:05 PM: Message edited by: JasonC ]

[mchlstrt]

I've noticed this enough myself to bump it. It seems worth a look by the B-Boys, if they haven't already.

strt

[IMP-22]

Well, when I played CMBO A Panzer IV could be easily taken out by a Sherman at just about any range. It might have been because of the less sophisticated way they modeled armor penetrations in CMBO. In CMBB at long range a Panzer IV can easily knock out a T-34, with the T-34s weaker gun not being able to penetrate the Panzer IV's armor, so it starts cowering.

I generally prefer to play singleplayer games mainly due to my terrible connection out here in the middle of nowhere, and I like to play as my own pace. So it's a bit frustrating when I set up a huge force of armor, ready for another Kursk, then the Russians start spinning around on the other side of the battlefield and my tanks destroy them with ease.

Another minor gripe I have with CMBB is how often turret hits are made. On certain tanks it is understandable. Like the T-34/85, for example. The armor on the front of its large turret makes a pretty big target, but a tank like the Panzer IV should not get front turret penetrations as often. It's turret provides a much smaller profile. You would think that if this had happened as often in the war as it did in CMBB, they would have increased the front turret armor on the Panzer IV instead of leaving it at 50 mm. Another example would be those little German half-tracks with the 20 mm cannons. The turrets on those things are very small, and don't provide a big target. Yet they get front turret penetrations all the time. The same thing happens to those little Russian armored cars. It seems like the AI in CMBB always aims high or something. I suppose this is a bit too much to ask, but it would have still been nice if the size of the turret on vehicles was taken into account when modeling where the shells from tanks hit.

[The Man who was Thursday]

I agree with everything in this thread.

[Zarquon]

As far as I know, the entire AI is hampered by the fact that it evaluates the situation each turn, but it cannot remember the position where an enemy unit was sighted as soon as LOS is broken. So a tank (or any other unit) might move into cover, loose LOS and rush forward again the next turn. Repeat until KIA. So a flanking maneuver is out of the question and the only remedy is to wait for The Great Engine Rewrite.

Still I would not bet that the problem will be addressed because making AI units remember things is easy but useless unless the AI is made into a *real* beast.

At the time, it acts and reacts instinctively, without a memory. Even the dumbest laboratory rat will learn from mistakes, sooner or later, because it posesses a memory. Not so the AI. I don't know all the different definitions of "intelligence", but to draw even the most simple conclusions from something, you have to be able to remember things first.

And only *then* comes the tricky part - make it recognize patterns (is this enemy tank I just saw tank part of a defense line of a counter-attack?), formulate plans (hey, I'll just do the gamey map-edge flanking trick!) etc. etc.

And when it is capable of this, go a bit further and make it actually *learn* from it's mistakes and successes.

We can only hope the next engine will not be *that* capable or we'll see BTS swallowed whole by the Pentagon - you'd have to wear a uniform to play CM II.

[Guest Sgt. Emren]

you'd have to wear a uniform to play CM II

Heck, I wear a uniform when playing CMBB!

[Ron]

I don't think either it is too much to ask. The Strategic AI is very weak and is unable to do even the simplest of things in a battle. The tank cower issue with the TacAI needs to be addressed also. It was apparent in CMBO at times, though very rare, but in CMBB happens more frequently.

The shortcomings of the Strategic AI have been well documented to BFC many times on this forum and their response has always been it isn't worth the amount time for only the marginal improvements that would result. In effect it is "good enough". I don't know as I am not a programmer :confused: I can relate to the people that are only able to play the computer, it IS a very unsatisfactory experience in situations; to see AI armour move mindlessly into the same killing zones, to see AI armour *sit* in the corner of the map throughout the battle, to see AI artillery assets go unused, to see AI forces leave defensive positions and begin moving with little aim, to see AI forces milling about for half the battle are a few examples.

To be fair there are situations where the AI performs adequately, but those seem to be the result of excellent scenario design rather than AI *brilliance*.

I doubt anything will change in the engine rewrite, though we can hope and be vocal enough so BFC takes notice and decides to invest their time. Until then it will be pbem/tcp play to get a rewarding experience and challenge.

Ron

[Engel]

AI is extremely hard to code. Most games that appear to have smart AI tend to do this by allowing the AI to cheat; the computer can do things faster, it has access to information denied from the human player, the AI is more accurate, it uses tactics that simply aren't available to human players, etc.

If BFC could code a non-cheating AI that is even half as good as a semi-competent human player, I don't think the goverment would let them use it in wargames

See the threads regarding AI scripting to see one solution to the StratAI incompetence problem; this is one AI feature that is feasible for the engine rewrite.

[MikeyD]

Don't paint a totally bleak picture of the AI.

I played a game last night against the AI - my one Hetzer and 3 Stug IIIs behind cover defending against an initial wave of wimpy SU-76s. I was sure I was going to dominate the battlefield no problem, but to my great surprise the SU-76s stood and fought... and wiped me out! I was genuinely stunned.

I've noticed the map seems to make ALL the difference. If the AI's armor starts going towards the flags by a playable route then it can put up quite a fight. If it starts moving by an undefensable route it has great difficulty backing up and starting over.

And another thing, what tactic SHOULD the AI have chosen when it spotted three King Tigers while travelling across open ground?

[CombinedArms]

Originally posted by Ron:

To be fair there are situations where the AI performs adequately, but those seem to be the result of excellent scenario design rather than AI *brilliance*.

Ron

I think a lot of the answer is here, at least until the engine redesign: if you're playing against the AI, play well designed scenarios, ones recommended by the designer or Scenario Depot reviewers for AI play. Most of these have maps and forces that have been shaped to allow the AI to perform at least adequately. Then give the AI an experience or force boost (or both). And expect to win, but enjoy the battle.

I tend to play battles against the AI when: 1) my pbem opponents haven't sent me a turn lately; and 2) I want to explore a historical situation or fictional problem treated in a given scenario; OR; 3) I want to try a scenario that comes highly recommended and looks like it would work vs. the AI. AI battles can actually work well to recreate certain historical situations--one's where one side won against seemingly impossible odds (AI dumbness actually helps one achieve the historical result) or battles where the attacker achieved (and needed) a significant degree of surprise. It's impossible to surprise a good human opponent in CM--we wouldn't be playing unless an attack was on the way.

I don't think QBs against the AI work very well in most instances--I only play them when I want to test out a tactic or piece of equipment or when I want to try to set up and solve a problem: e.g. how to fight against Panthers with T34/76s. In that case, I'm generally giving the AI the tactical or equipment advantage and trying to find a way to overcome it. If you give the AI a significant tactical-equipment disadvantage, it generally isn't going to succeed. You're asking to much from it--for example, with those Tigers on a ridgeline. In general, I think pure armor QB's vs. the AI can be quite entertaining if you let the AI have the superior tank killers.

So, next time, try giving the AI the platoon of crack Tigers on the ridgeline (remember that the AI is better at defense) and YOU trying attacking with T34/76s. See how you do--oh, and don't forget to report back.

[ March 31, 2003, 01:04 PM: Message edited by: CombinedArms ]

[JasonC]

The problem of the Tac AI must not be confused with the general problem of the strategic AI. The Tac AI problem hurts even human vs. human games. It is not avoidable, or acceptable. Weakness of the strategic AI can be countered by playing humans, or by giving the computer high odds. Weaknesses of the Tac AI wreck human vs human games, and make some tactical situations so unrealistic as to be unplayable. A quest for perfection in the strategic AI may not be used as an excuse to avoid fixing the Tac AI when it is broken.

It is asked, what should the AI do when it sees 3 Tigers in the open? Every AFV *targeted* by one of the Tigers, with the Tiger targeting it facing its turret within 30 degrees of that AFV, should fire one round and should then pop smoke if avalable and head for cover, reversing if necessary. Other than getting off one round, right now this happens *with all* AFVs, regardless of targeting or facing. That is the problem.

AFVs that are *not* targeted by one of the Tigers, or that the Tigers are *not* facing, should *continue their existing orders*. Which may mean advancing, it may mean shooting, it may mean heading for cover but forward rather than back.

The *combination* of these behaviors allows a *group* of inferior AFVs to fight against superior AFVs. It does not make it easy. They can be killed in turn. But if they team up, managed to close, get flank shots, etc, they can shoot at the superior AFVs and potentially hurt them (with gun damage etc "hail fire", or close penetrations etc).

What needs to be added is most of all a checking subroutine before the "cower" flag is set. That checking routine, in pseudo code, asks -

Am I targeted by a superior AFV?

If not, no cower

Is the targeting AFV facing within 30 degrees of my present location?

If not, no cower

am I ready to fire?

If so, no cower

Notice, the last will change the second after firing, and lead to cower during reloading, if targeted and faced by a superior AFV.

All tanks will try to turn and get off one shot. After that one shot, they will cower if targeted and faced. They will not cower if not targeted, or if they have "the drop" on the enemy AFV (meaning it is facing the wrong way).

If 30 degrees seems too narrow, make it 45 degrees. The intention is to allow *a pair* of tanks seperated by an angle of 90-120 degrees, to have *at least one* of the pair *not* cower, when facing a single superior opponent.

The basic method of dealing with superior AFVs is *not* "pop smoke and run away". It is *teamwork* and *hail fire*. Yes, those immediately threatened *should* "pop smoke and run away". But *while* playing their part in the "teamwork and hail fire" SOP.

It still takes skill in the orders given to the team of tanks to use teamwork and hail fire *successfully*. But it is *possible* to perform it successfully, if the Tac AI doesn't *override* player given orders with its idiot "run away" SOP.

The strategic AI does not need to be improved to benefit from this change. It already gives orders to its tanks to move about. Those orders just won't be *cancelled* and replaced by non-sensical, individual, disorganized panic responses, regardless of numbers, who has the drop on whom, etc.

[ March 31, 2003, 01:05 PM: Message edited by: JasonC ]

[Hensworth]

Anyone who fails to grasp what JasonC is banging on about, play the Jaegermeister scenario as the Russians.

[tar]

I'm not so sure about the real viability of the "if not targetted" condition.

That has the potential to lead to complaints where instead of (say) two out of three tanks ambushed by a single superior vehicle reversing out of danger, they will each wait, in turn, until they actually get targetted before trying to hide.

Then we'll get a thread about how stupid the TacAI is, waiting too long to run and hide, when anybody can see that they are doomed if they just sit there.

I would think that the most effective solution would be to include something like the TacOps style SOPs into CM. The real problem is that the reasonable behavior depends a lot more on the situation than one can reasonably expect the TacAI to figure out on its own.

In some ways, this is even more true when it is executing the human's plans, because with the StratAI in control, there is at least some chance, small as it may be, that the TacAI will know what the strategic plan is. With player control, there is no way for the TacAI to guess whether pressing the attack or force preservation is more important.

[von Reichardt]

Originally posted by Zarquon:

As far as I know, the entire AI is hampered by the fact that it evaluates the situation each turn.

I agree. Another thing is that the AI don't seem to have any human pre-programmed attack strategys (by the programmers) to adapt, it seems that the AI just setup on line and then charge. A couple of good human programmed strategies (some 30 or 40 different ones) which could sometimes change in the middle of a scenario would make things more unpredictable and interessting.

Otherwise I agree to most of the things in this tree

[White Phosphorus]

Originally posted by IMP-22:

Another minor gripe I have with CMBB is how often turret hits are made. On certain tanks it is understandable. Like the T-34/85, for example. The armor on the front of its large turret makes a pretty big target, but a tank like the Panzer IV should not get front turret penetrations as often. It's turret provides a much smaller profile. You would think that if this had happened as often in the war as it did in CMBB, they would have increased the front turret armor on the Panzer IV instead of leaving it at 50 mm. Another example would be those little German half-tracks with the 20 mm cannons. The turrets on those things are very small, and don't provide a big target. Yet they get front turret penetrations all the time. The same thing happens to those little Russian armored cars. It seems like the AI in CMBB always aims high or something. I suppose this is a bit too much to ask, but it would have still been nice if the size of the turret on vehicles was taken into account when modeling where the shells from tanks hit.

This has been beaten to death by JasonC and others in the numerous "early T-34s are are too weak" threads. For some reason, the game treats all turret fronts as 30% of the hull exposure.

[JasonC]

Tar said "I'm not so sure about the real viability of the "if not targetted" condition. That has the potential to lead to complaints"

No one will complain about their units continuing to follow the orders they themselves gave to them a tenth as much as they (and I) complain about their orders being cancelled to be replaced by nonsensical ones that make fighting impossible.

If the tank isn't targeted it does not spontaneously combust or commit suicide. It follows its orders. If those orders wind up proving suicidal, it is the fault of the commander who gave the orders.

If those orders get it behind terrain cover, or onto the flank of the enemy, or take a side shot successfully, or contribute to "hail fire" that disables the enemy, or close the range - then the credit is to the player who gave the order that wound up being sensible.

If, however, a player gives orders that would get a tank behind terrain cover long before the enemy had any chance to shoot - or that would take out the enemy with a flank shot before being targeted - but the tank refuses to carry out that order because it instantly panics the minute LOS is established - then the fault is not the player's. It is the *busted* Tac AIs.

[flammenwerfer]

Bump

Any comment from BFC?

[tar]

Tar said "I'm not so sure about the real viability of the "if not targetted" condition. That has the potential to lead to complaints"

JasonC: No one will complain about their units continuing to follow the orders they themselves gave to them a tenth as much as they (and I) complain about their orders being cancelled to be replaced by nonsensical ones that make fighting impossible.

Players will complain about any AI behavior.

I was actually taking care to isolate this one factor in JasonC's formula. I could see the other issues (such as facing -- or perhaps more importantly gun facing) figuring into this, but not the question of being targetted. The advantage of gun facing is that it correlates with the potential threat, and I find that knowing that you are actually being targetted when no firing it going on is a bit questionable from a realism standpoint.

-Tom.

[JasonC]

Fair enough. If the facing "threat zone" is kept relatively narrow, within 30 degrees, the targeting portion of the test could be skipped, I suppose. If the threatened tanks all take one shot before reversing, until some can get to a flank, it would still be possible to fight.

Right now, it isn't.

[Yaba]

Just bumping this beacause it addresses my situation in my last post (suicidal tank commanders).

Does anyone know if this has been addressed in the new 1.3 patch?

Cheers,

[walpurgis nacht]

Originally posted by JasonC:

No, IMP-22 is not asking too much. Tank cower behavior is extreme in CMBB, and it is broken.

The main problem is that tank cower behavior does not take into account the facing of the superior "scary" vehicle. This needs to be changed.

At present, it hurts not only the clueless AI but even the best possible human directed play. The Tac AI fumbles the "hand off" for the last 5 seconds.

If you expertly maneuver a T-34/76 with tungsten into a rear shot on a Tiger I at 100m range, with the Tiger facing completely the wrong way, the T-34 will often refuse to take the shot, and instead reverse the instant LOS is established. This is as realistic as a bucket of warm spit.

Jason, you have written this many times . . .saying the "cowering" is "broken", etc. I have not had this problem . . . almost never in fact, and I attack tigers and panthers with t34/76 (late model with tungsten) all time. Yes, the t34/76 will "cower", but only if you're doing it incorrectly. You say you are "expertly" manuevering your t34s in to kill the tigers. Please elaborate with greater detail . . . you are doing something wrong if they cower more than 5% of the time.

The key is to "fast" move . . .never "hunt" or "move to contact". If you are behind a hill, 110 meters about to pop up and shoot the tiger in tail . . . "fast" move him to 10-20 meters behind the tiger, not 100 . . .even with tungsten, that way you "force" him to engage . . and I promise you at that range the t34 will not cower. Obviously this is suicide unless you have another decoy t34 to keep his turret busy in the opposite direction for that 20 seconds it takes you to get there, so coordination and timing is key. I would be glad to do a small tcp/ip demonstration if you can't get it to work.

Hope this helps.

[JasonC]

"You say you are "expertly" manuevering your t34s in to kill the tigers. Please elaborate with greater detail"

I am refering to the whole approach to close range from opposite sides of the map, use of available cover, distraction and misleading, predicting where the Tiger will go. And maneuvering a "shooter" T-34 with T ammo behind cover (low ground, large woods) to approach the Tiger from a blind area at close range.

Specifically, a Tiger is positioned at a gap between two bodies of woods, hunting forward toward the distraction force. It is prevented from charging too rapidly by the threat of side shots from two directions if it clears the trees on either side immediately (giving up its "keyholing"). The distraction force is staying alive by changing positions, and making the Tiger crawl forward to re-establish LOS.

Meanwhile, the shooter T-34 has done an end run around the left side woods the Tiger is "keyholing" between. It runs along the far side of those woods, approaching to within 80m of the Tiger with full woods between, blocking LOS. Its last move is to fast move directly behind the gap between the two bodies of trees the Tiger is keyholed between.

That will put it right on the Tiger's ass at 40m range, with the Tiger's turret 150 degrees out of position (roughly 30 degrees left of its hull, searching for part of the distraction force). The Tiger has no shot during the whole approach. LOS will first be established between 80m and 60m, about 135 degrees to the right rear of the Tiger. It will have only a few seconds to turn.

The T-34 is in radio command, +1 morale tank commander, has regular quality, is unphased in all respects, and has 4 rounds of APCR. The Tiger is fully IDed (by the distraction force elements, which have had it is sight for several minutes). It has a vehicle covered arc to the right front, about 90 degrees width, that places its present turret facing exactly as required to shoot the Tiger when LOS is established.

Resolution phase - the T-34 crawls to the edge of LOS. It then halts. The Tiger acquires and a targeting line runs from the Tiger to the T-34. Its turret begins slowly turning, but is still over 120 degrees off proper facing. The T-34's turret is perfectly aligned. Where it halted (against orders), the T-34 has a side angle to the Tiger rear of around 30 degrees.

The T-34 does not fire. Instead it reverses, wasting time to rotate slightly. A straight reverse would break LOS instantly, but it doesn't. Instead it reverses along a line closer to 180 degrees away from the Tiger, which tends to keep LOS. The Tiger turret rotates a total of 150 degrees without the T-34 firing once. The Tiger fires once. The T-34 burns.

Ahistorical horsefeathers and side splitting hysterical. Tank cower is broken. Fix it please.