CMBB-What it REALLY needs!(better Strat AI)

[Panzer76]

Hi there!

Just played a rather big scenario that comes with the game, "All or Nothing". I played as the Germans (defenders) and real fast I saw what really needs improving, the AI. It really struggels with big maps and lots of units and objectives far from each other. The thing that happend here was just pathetic, the Brits took the first VC and then just did nothing.

And there is more, the AI drove the Tanks and vehicles all over the place, into forests, into each other, back and forth, without any logical reason what so ever! The inf wasn't spared either, they came in trickles and was guarding nothing in the middel of nowhere, just stood there.

The result was a extremly boring game, and just silly. 10% of the British forces saw any action, the majority was just doing stupid stuff all over the map.

I understand that there is a TacAI and a StrAI, but it is obvious that on (ecspecially) big maps it plainly doesn't work. And this should be a worry as the maps in the east can be and usually will be gigger than those in CMBO. Usually I just give the Computer way more forces than I have, but whats the use when it doesn't even use them?!

BTS: Have you improved the AI for CMBB? Please say so, or else it will almost be a multiplayer game only.

[ April 06, 2002, 11:21 PM: Message edited by: Panzer76 ]

[Slapdragon]

Can you give a tree or a diagram showing how you would improve the tacAI? Most people feel the AI is one of the most advanced elements of CM -- especially since AI programming is so difficult.

I am sure BTS is working on the basic AI, but I would also like to see your idea for making a better one.

[Lars]

What CM really needs is a preview mode that would show the best move instead of the one I plotted.

[aka_tom_w]

First Off I know NOTHING about programing or AI

so to comment here might risk appearing foolish.

since that has never stopped me before I won't let it stop me now

Would it be possible to program the AI to be more interested in destroying enemy forces then concentrating on holding objective flags?

It seems perhaps too simple to just say:

i) Scout enemy units,

ii)use max available firepower and

iii)destroy all enemy units.

Again that seems WAY too simplistic but maybe the AI could have a little more "killer instinct" written in to attempt to destroy the opposing force.

That said, the AI will defend very well. If you set up the units in a good defensive posture and ask the AI to defend it does defend admirably.

-tom w

[ April 04, 2002, 11:35 AM: Message edited by: aka_tom_w ]

[Slapdragon]

Originally posted by aka_tom_w:

First Off I know NOTHING about programing or AI

so to comment here might risk appearing foolish.

since that has never stopped me before I won't let it stop me now

Would it be possible to program the AI to be more interested in destroying enemy forces then concentrating on holding objective flags?

It seems perhaps too simple to just say:

i) Scout enemy units,

ii)use max available firepower and

iii)destroy all enemy units.

Again that seems WAY too simplistic but maybe the AI could have a little more "killer instinct" written in to attempt to destroy the opposing force.

That said, the AI will defend very well. If you set up the units in a good defensive posture and ask the AI to defend it does defend admirably.

-tom w

Tom, the usual problem with these discussions (many have been done before) is everyone wants a "better" AI but no one can turn their ideas into some form of math that game designers can code into small units (like, how do I react to x unit) or how these little AIs feed into the big idea.

But still, you are basically on the right track. Defining the problem if the first issue needed to improve AI. It is not the only issue, but it is the start.

[Holien]

Hi Panzer76

Why the sad face?

You said: -

"BTS: Have you improved the AI for CMBB? Please say so, or else it will almost be a multiplayer game only. "

I know that the game can be played by yourself and that is a challenge to begin with. But what is wrong with a multiplayer game only?

If you think about it we can not possibly hope to mimic the devious nature of human players in CM. So accept the limitations and choose some fellow players to get your fix on.

Sure that might not be ideal but I can assure you it is far better than when there were no computers and your mates lived some distance away and your mum kept wanting the spare room floor back and crushing your airfix men and tanks.

It is better than trying to load a computer game from a crappy old tape drive and waiting 15 minutes for it to load only to then find that it had failed to load on the 48k Spectrum you had set up with the TV.

I like Snappy await the day some nice young person comes along and develops a new way of doing the AI that is so blindingly obvious no-one sees it at the moment.

Until that day play humans to be stretched further than you ever want to be stretched.

H

Edited to be PC and stop her in doors complaining

[ April 04, 2002, 01:41 PM: Message edited by: Holien ]

[Panzer76]

Originally posted by Holien:

Hi Panzer76

Why the sad face?

Of course, it will be a LONg time untill the AI will be as good and unpredictable as a human, and playing against a human is certainly fun, but sometimes its ok to just play the CPU, and then it would be nice if it could give me a rael match, and dont do extremly silly stuff.

[Redwolf]

The real deal would be a computed opponent that is replacable by the user, in the form of a shared library (*.DLL files in windows). Of corse, both sides should optionally be controled that way so that two guys can have their AIs shoot it out against each other. Most advanced: Mixed AI and player play, AI plots, Player corrects and hits GO - get rid of much boring infantry fineplotting.

Instead of trying to fix the TacAI endlessly we need a better mix of the existing TacAI, but more tuning commands (like cover arc in CMBB), SOPs like TacOps has, of even boolean decision trees like Steel Beasts is supposed to have (reminds me, time to get that game).

[aka_tom_w]

So I was curious....

hm.. "Boolean Decision Trees" I wonder????

so I look it up in a web search...

this page comes up

http://epubs.siam.org/sam-bin/dbq/article/31010

OK this stuff looks SERIOUSLY complicated

maybe we need some of these Boolean Decision Trees

in CMBB????

Yeah Right !?

can anyone here make any sense out of this article

its seems VERY academic and geeky to me?

-tom w

Average-Case Lower Bounds for Noisy Boolean Decision Trees

William Evans, Nicholas Pippenger

Abstract. We present a new method for deriving lower bounds to the expected number of queries made by noisy decision trees computing Boolean functions. The new

method has the feature that expectations are taken with respect to a uniformly distributed random input, as well as with respect to the random noise, thus yielding stronger

lower bounds. It also applies to many more functions than do previous results. The method yields a simple proof of the result (previously established by Reischuk and

Schmeltz) that almost all Boolean functions of n arguments require $\Me(n \log n)$ queries, and strengthens this bound from the worst-case over inputs to the average over

inputs. The method also yields bounds for specific Boolean functions in terms of their spectra (their Fourier transforms). The simplest instance of this spectral bound yields

the result (previously established by Feige, Peleg, Raghavan, and Upfal) that the parity function of n arguments requires $\Me(n \log n)$ queries and again strengthens this

bound from the worst-case over inputs to the average over inputs. In its full generality, the spectral bound applies to the "highly resilient" functions introduced by Chor,

Friedman, Goldreich, Hastad, Rudich, and Smolensky, and it yields nonlinear lower bounds whenever the resiliency is asymptotic to the number of arguments.

Key words. fault-tolerance, reliability, noisy computation, error-correction

[Redwolf]

Tom,

Originally posted by aka_tom_w:

hm.. "Boolean Decision Trees" I wonder????

I didn't mean to refer to any existing term.

It's just TacOps-like SOP extended by "if this and that is true at this point, the follow this path and do this action, if the, follow the other path and do this instead". You build a tree where each leave is a command set and each brnach point is a question expressed in boolean logic (and, or, not).

As I said, I don't even have Steel Beast, it's just what I've been told it has. Anyone can shed some light on it?

[Gpig]

Awwww, COME ON Panzer76!

I've played All or Nothing against the A.I.. But I played as the Allies.

It was an AWESOME battle. It was hard enough on me (as the British AND a human) getting my attack to work. I have NO illusions about the A.I.'s abilities. But to have to deal with three river crossings and multiple avenues for attack. Plus a constant flood of reinforcements . . . I'd be amazed if ANY computer could deal with the logistical nightmare of that scenario (for the allied side).

I truly think your expectations of the A.I. are TOO HIGH.

Part of your experience is due to the fact that you have yet to learn the best ways to play the A.I.

There are few "best" ways to play the A.I..

But one of the wrong ways (if you want a poor showing by the A.I.) is to let it attack you.

Some scenarios are custom made to allow for a pretty good showing for the A.I. on attack. They usually recommend how to adjust the balance for the A.I. in their briefings. They use flags, and exit zones, and friendly/non-friendly map edges to help coach the A.I. into a workable attack. Check 'em out.

But overall, Let the A.I. defend.

And DO NOT play the A.I. (where you are the germans) in ALL OR NOTHING. (Guess you already know that, though.)

Gpig

[ April 04, 2002, 02:28 PM: Message edited by: Gpig ]

[benpark]

This may be a bit simplistic, but would there be a method within the scenario design section that could specify a basic path of attack/ assault/ probe? It may make playing games that you designed predictable against the AI, but for designing historical situations-it would be nice.

I realize that flags can do some of this, and this may be too much on the strat side. This may be countered by grouping units and assigning them to a path of attack.

I find the AI fairly hopeless in big scenarios, and sometimes concerning bridge crossings.

Defend Could be done roughly the same way as the traditional flag method, maybe with a few tweeks, as the AI gives a pretty good fight on D.

[Smiler]

Originally posted by Slapdragon:

...everyone wants a "better" AI but no one can turn their ideas into some form of math that game designers can code into small units (like, how do I react to x unit) or how these little AIs feed into the big idea.

Here, here. There's nothing wrong with requesting better AI in the future, but in reality I think the CMBO AI is pretty impressive. That is, it's good AI, even though it's bad at some things.

To get a feel for how difficult AI is to get right, imagine that you're playing CMBO, but you can't see the battlefield (including units). Imagine trying to plan a strategy when your input is "you have a tank at (10,10), and a rifle platoon at (10,12). There are woods at (11,10), (12,10), (13,10), etc, etc". It's not easy.

The human mind's ability to process visual information at a high level (e.g., where all the units are relative to each other) is extremely powerful and something that computers simply can not do easily. Just consider how many games get optimal pathfinding correct (very few), which on the surface seems like a trivial problem to solve.

[Pvt. Ryan]

IIRC, CMBO was designed with battles of a maximum of 1,500 points and therefore smaller maps. I don't think the AI was changed when the points maximum was increased, and as a result, the AI can't deal with the larger map area as well as it could when it was confined to a smaller area.

[Andrew H.]

Originally posted by Gpig:

Awwww, COME ON Panzer76!

I've played All or Nothing against the A.I.. But I played as the Allies.

It was an AWESOME battle. It was hard enough on me (as the British AND a human) getting my attack to work. I have NO illusions about the A.I.'s abilities. But to have to deal with three river crossings and multiple avenues for attack. Plus a constant flood of reinforcements . . . I'd be amazed if ANY computer could deal with the logistical nightmare of that scenario (for the allied side).

Gpig

That was sort of my experience, too. In fact, when I read Panzer's description of the AI attack:

And there is more, the AI drove the Tanks and vehicles all over the place, into forests, into each other, back and forth, without any logical reason what so ever! The inf wasn't spared either, they came in trickles and was guarding nothing in the middel of nowhere, just stood there.

my first thought was that this described *my* attack against the AI in All or Nothing.

[Panzer76]

Originally posted by Andrew Hedges:

my first thought was that this described *my* attack against the AI in All or Nothing.

Wanna play a multiplayer against me?!

[Panzer76]

Originally posted by Pvt. Ryan:

IIRC, CMBO was designed with battles of a maximum of 1,500 points and therefore smaller maps. I don't think the AI was changed when the points maximum was increased, and as a result, the AI can't deal with the larger map area as well as it could when it was confined to a smaller area.

Yes, but in CMBB it's not like that anymore. And what has been done to ensure that the AI has developed alongside the increased size of the battles?

[Panzer76]

Originally posted by Smiler:

The human mind's ability to process visual information at a high level (e.g., where all the units are relative to each other) is extremely powerful and something that computers simply can not do easily. Just consider how many games get optimal pathfinding correct (very few), which on the surface seems like a trivial problem to solve.

No one said that it was easy, it's extremly difficult. But with ethe sizes of the battles getting bigger in CMBB, something has to been done.

[Fred]

Well,

the tactical AI in CMBO is very well done, and better than anything else I saw in other games.

What REALLY needs an enhancement is the strategic and operational AI (and I have the impression that this AI was better in the first versions of CMBO).

The argument, 'that it is not possible to make a good AI with today techniques' does not hit. Just take a look at the already implemented TAC AI!

And if all programmers stop working on AI enhancements, because 'it will not be better', well, then, indeed, nothing will ever happen, like a self-fulfilling prophecy.

To be honest, I would prefer a better AI anytime over eye candy in CM.

What are the main problems of the AI?

1. The AI has never heard about platoon cohesion. Usually the platoon HQs move like independent units. Very bad, because that let all AI units move and react slow, and they do not get the HQ bonuses.

All the AI has to do is keep the platoons together.

With the new rule for tank platoons, this should be fixed.

2. Leading with tanks.

There is almost no situation in which tanks should lead an attack.Period.

3. AI Tanks never ever button up, once buttoned. Even in multi-battle opearations, AI tanks stay buttoned from battle to battle.

This should be fixed in CMBB.

4. No variation in movement logic.

All AI tanks always want to go to exactly the same spot at the same time, because they use the same routines for their movement plot; traffic jams and tanks 'dancing around' are the result. Turkey shooting for any AT assets. This happens when the AI has quite some vehicles.

A simple variation, while using the same routines (Tank A moves and turns here; Tank B moves and turns 5 meters left of Tank A, etc.) would help.

5. Not assigning troops to the battle.

Often after a battle one could see, that up to one third of the AI forces never moved or only moved some meters from their start position; no shot fired, not in the vicinity of any flag, etc.

So I am not asking for an AI revolution, but for an evolution.

Most people are often playing the AI, I would even say that the majority plays the AI on a regular basis.

And do not forget, even in a human vs. human game, after hitting GO, the AI takes over...

Fred

[KNac]

Well said Fred. I think that the TacAI is very good, but Strat AI, is not very good in large battles and with large forces.

IE. I´ve done an operation (I´m still testing) that is huge, it really push CMBO to the limits with a full strengh inf regiment with lost of supports assets, a part of other problems, the heavier problem was with the AI, in the first battle the US in on defense against the enemy german regiment, because I must position all units of the germans yet, I picked up allied side to see how it played against AI.

First of all, I must tweak the scenario to assault (when it should be advance) because the AI didn´t move (I haven´t tried destroy yet). When I did the AI played just in a stupid way, with a full regiment it only moved some of its assault guns which were destroyed easy by my AT guns and TD and it didn´t use its infantry (with a full regiment lol). Also you have noted that it leaded with tanks. I´m still testing but there is some problems within thje CMBO that cannot make the type of operation I want.

I know it is a bit off-topic but I will coment anyway. I haven´t written the brief but it should be something like this:

It takes on June 44 (it is fictional but) before D-Day when the allies are pushing towards Caen, etc. The fictional scenario is this: the germans have exploited a gap in the flank of an advancing us armored division and its corps (or army), a lot of units haven´t entered in fight yet (because of that base level is green and are full strengh) a kampfgruppe formed by an inf regiment with lots of support units (and artillery), and an assault gun coy have been formed to exploit the gap. When a inf bn was trying to link between the division and its corps receibed a radio comunucation that a large german force was going towards their position. Well the US mission in the first battle is to hold the ridge they are defending and stop enemy advance until reinforcements, after that the US receive LOTS of reinforcement and must counter atack and destroy the enmy attemp to break the link (and the main supply route) and put that division out of supplies creating a pocket that could destroy that division.for the germans after reaching the ridgeline, it would receive the rest of the kampgruppe assets (2 panzerjäger coy of different types) to defend the ridgeline and if possible expand the gap. I think is a nice scenario that involves low level strategy and nice tactical action. I don´t demand the AI to be so inteligent but at least it could move units an other lack on the CMBO engine you can note in this operation is the option of define the caracter of each battle, I mean, now you only can use 3 general status for the whole operation: advance, assault and destroy 8and if u pick defender: defend) but what about if you want to make an scenario like i described, it should be 1&2 battle assault (or take x flags if they could be used, dynamic flags that after that scenario dissapeir) and before that it should be defend. and for the US side it sould be defend (1st and maybe 2nd battle) and before atack (or counter-atack that is the same). I know is a bit off-topic but it seems a ncie option and easy to add. Anyway the op I have done is more to play against human, if someone wanna play it when is finished (but it is huge!)

[ April 05, 2002, 06:30 AM: Message edited by: KNac ]

[WWB]

There is a reason the TacAI does much better than the StratAI: it has alot less ground to cover. Input is generally limited, mainly being target and los information.

On the other hand, the StratAI has to work with every conceivable setup, and is really being overburdened by the newer, more complex scenarios. It does a pretty decent job on a simpler map in the defensive, but really is not up to parsing a huge, complex map (such as all or nothing) and successfully making a convincing attack. That is a truly difficult proposition to handle without:

a) Cheeting

or

Being given a specific plan

That said, I think some things could be improved without putting in undue effort. Principally giving the StratAI an idea of what formations are as well as giving it a hint of which units should not be pushed with the front line troops.

WWB

[Fred]

Originally posted by wwb_99:

There is a reason the TacAI does much better than the StratAI: it has alot less ground to cover. Input is generally limited, mainly being target and los information.

On the other hand, the StratAI has to work with every conceivable setup, and is really being overburdened by the newer, more complex scenarios. It does a pretty decent job on a simpler map in the defensive, but really is not up to parsing a huge, complex map (such as all or nothing) and successfully making a convincing attack. That is a truly difficult proposition to handle without:

a) Cheeting

or

Being given a specific plan

That said, I think some things could be improved without putting in undue effort. Principally giving the StratAI an idea of what formations are as well as giving it a hint of which units should not be pushed with the front line troops.

WWB

Exactly.

Some (realative) small changes, and suddenly the AI keeps unit cohesion and does not lead with tanks.

A vast improvement without to much effort.

And, btw, creating a challenging AI is very much fun; did this myself and I stopped once I was beaten by the AI 50+% of the time...

(only a very small game, man-to-man...computer version of Strike Team Alpha...)

But it is possible...if even I got some decent AI

Fred

[Slappy]

Fred's assessment of the largest AI failings are dead on. In fact, I was ready to post somthing much like it until I got down there. There is only one real addition that I have: Yery bold Spotter Use. In my opinion, the AI puts too high a premium on direct LOS for artillery. This puts the Spotters in the front lines and generally means that only 50% of AI rounds are used.

As for AI improvements, they are really tough. Even correcting these things requires more work than you think.

Example: Platoon cohesion. You introduce a factor for this, or up an existing factor to make the AI place higher value on keeping units in command. (I have no idea what the AI algorithim does, but my guess is that it tries to optomize some value that approximates victory points). Sounds easy, but there are many times that you want units to be out of command, at least for short periods of time (withdrawls, spread out defence, splittng to scout terrain). Terrain, HQ bonuses and troop quality can have a huge impact on the distance for 'in command'. This becomes, in association with an 'in turn' AI very difficult to calculate. What if a platoon loses cohesion due to enemy fire. How should it get back in range? Should some of the squads re-cross the ground under fire? Which ones? Can it form up somewhere else forward or to the rear? What delay in advance should be acceptable to this? As it turns out, weighting cohesion more highly can acually exascerbate the 'not committing all units to the battle' problem. Moving into contact increases the risks of loss of cohesion and therefore makes staying in one place a relatively better option. Contrary to popular belief, there is no way to really 'program' complex behaviors into the AI. You have to set up goal weightings to this and watch the AI to see what happens and then tweak the result, sometimes with unintended consequences.

[KNac]

This is an example than a computer will never be smarter than a human the best is playing against other humans

[WWB]

Originally posted by Fred:

</font><blockquote>quote:</font><hr />Originally posted by wwb_99:

There is a reason the TacAI does much better than the StratAI: it has alot less ground to cover. Input is generally limited, mainly being target and los information.

On the other hand, the StratAI has to work with every conceivable setup, and is really being overburdened by the newer, more complex scenarios. It does a pretty decent job on a simpler map in the defensive, but really is not up to parsing a huge, complex map (such as all or nothing) and successfully making a convincing attack. That is a truly difficult proposition to handle without:

a) Cheeting

or

Being given a specific plan

That said, I think some things could be improved without putting in undue effort. Principally giving the StratAI an idea of what formations are as well as giving it a hint of which units should not be pushed with the front line troops.

WWB

Exactly.

Some (realative) small changes, and suddenly the AI keeps unit cohesion and does not lead with tanks.

A vast improvement without to much effort.

And, btw, creating a challenging AI is very much fun; did this myself and I stopped once I was beaten by the AI 50+% of the time...

(only a very small game, man-to-man...computer version of Strike Team Alpha...)

But it is possible...if even I got some decent AI

Fred</font>