Is the strategic AI the bottleneck of CM game diversity?

[Leonidas]

Originally posted by Hurricane:

What I would like to see is an AI which is dynamic, i.e. can adapt to the situation and which learns from its mistakes. I´d love to see an AI built on a neural network, not on a few thousand lines of if-elses. Throw in a simulated memory, and the AI could become a very hard enemy to beat. I know it is nest to impossible to program such an AI, but maybe in CM 10?

Everyone would love to see this. There are lots of researchers working on these things full time. If it happens, I'm sure the initial applications will be in areas much more lucrative than computer games. But from what I can tell, nobody is anywhere close to dramatically improving AI. In fact, I don't think anyone even has a very good theory as to how AI might be dramatically improved.

[SuperSlug]

I'd like to see the AI implemented as a plug-in of some type. So another developer could implement his own improved version.

This has been done in the chess world for years now. I don't see why you couldn't extend it to CM.

Wouldn't it be neat to see a CM AI tournament?

As for more detailed orders. The best way to implement something like this is to give the player a scripting language so he could write his own generic order routines. At game time you would just select one of the existing CM orders or one of your scripts.

So to get you zook team to hide then fire twice and then deevee.

(HIDE)

(ENEMY AFV DISTANCE <= 150)

(R = FIRE ATW)

(IF R = 0)

{R = FIRE ATW)

(/IF)

(WITHDRAW = 100)

(/HIDE)

You get the idea...

[This message has been edited by SuperSlug (edited 03-16-2001).]

[Leonidas]

Originally posted by SuperSlug:

As for more detailed orders. The best way to implement something like this is to give the player a scripting language so he could write his own generic order routines. At game time you would just select one of the existing CM orders or one of your scripts.

That would be too detailed for me, and I thought I was extreme on the detail issue.

I want more detailed orders because I think that real soliders get pretty detailed orders. That is, they know whether they are free to use ammo or whether they should be saving it. They know whether they are trying to avoid enemy detection or whether they should engage any targets they see. They know whether they should keep their AT gun hidden until they see armor, or whether the brass has concluded that there is no armor in the area, and they should load HE and fire at infantry. The tanks know whether enemy armor is still considered a threat in the area, or whether the area is considered clear of enemy armor, so they should engage infantry.

But even I wouldn't want to go beyond that, to writing actual scripts.

[SuperSlug]

You are missing the point.

Scripts are portable/downloadable. A website has a list of detailed orders. You download the ones you think you might like to use and put them in a folder. At game time. One of the menu options for your troops/vehicles orders is [select detailed order]. You pick this and then are given a list of the detailed order scripts you have. Pick the one you want and voila!!! Of course you could write you own or tweek the existing scripts. At turn time the engine executes the script you chose.

So no longer will the BTS developers be nagged for more order types. You can create your own.

Cheers,

[This message has been edited by SuperSlug (edited 03-16-2001).]

[This message has been edited by SuperSlug (edited 03-16-2001).]

[Guest Andrew Hedges]

Originally posted by SuperSlug:

I'd like to see the AI implemented as a plug-in of some type. So another developer could implement his own improved version.

This has been done in the chess world for years now. I don't see why you couldn't extend it to CM.

Wouldn't it be neat to see a CM AI tournament?

I don't think it's quite right to compare CM's AI (and I think you're talking about the Strat AI here, not the Tac AI) to the AI's used by chess programs. The rules for chess are widely known and much less complicated than the rules for CM; also, there is no randomness for chess.

This would also mean that CM would be required to provide other developers with detailed explanations of how everything in CM works, from fuzzy logic routines to chance of suppression at different firepower rates. I don't think that they are interested in giving away their intellectual property like this.

Although I wouldn't mind some more variation in the StratAI's approach.

Keep in mind, too, that the success of chess programs such as Big Blue is not due so much to improvements in AI (although that was an original reason IBM was interested) as it was due to the fact that programmers, etc., were able to successfully reduce the game of chess to a large number of algorithms.

------------------

WOOT!

[Leonidas]

Originally posted by SuperSlug:

You are missing the point.

I think I understand the idea. And there are lots of games where this kind of approach would be interesting, though implementation can be problematic. I've made a detailed proposal for multiplayer Thief that would involve what you're talking about, where the script files detail guard patrol routes and behavior.

But scripts can have unintended consequences. Tribes had an extensive scripting system, but it was a little too powerful, and became a bit of a problem. The only way to be good at the game was to get good scripts, which meant you had to spend hours combing through web sites or personally learning the scripting language. Neither of those activities was much fun, so Tribes quit attracting casual gamers. I hope they eliminate it or cut it way back in the sequel.

[SuperSlug]

Originally posted by Andrew Hedges:

I don't think it's quite right to compare CM's AI (and I think you're talking about the Strat AI here, not the Tac AI) to the AI's used by chess programs. The rules for chess are widely known and much less complicated than the rules for CM; also, there is no randomness for chess.

Yes I was referring to the Strat AI. Sorry for the confusion.

I think that the rules governing troop movement and spotting could probably be distilled down quite a bit without giving away intellectual property.

A unit is a stateful object.

Has it been spotted?

What does it see?

Is it in cover?

etc etc.

Just these few items would give you lots of input into how the unit should behave.

The very fact that BTS has been able to come up with a fairly good AI indicates that it is quite do-able.

If the Strat AI is not opened up through and API or something, it will continue to evolve very slowly and even the novice player will have no trouble winning against it for the forseable future.

[ADAman]

From Terence: "I like to play against the AI, not because it is always the most challenging opponent, but because it is always ready to play, never argues with me and does its turns really fast"

____________________________________________

For me, I'd also add it doesn't gloat when I pull a bonehead maneuver...

------------------

Air Defense: Shoot 'em down, sort 'em out on the ground (AKA - if it flies, it dies)

[jasoncawley@ameritech.net]

Two bad ideas on this thread, and one on which I have less of an opinion.

Scripts are a horrible idea. I say that as one who played the spread-sheet like 4X space game "stars" for several years and wrote most of its strategy guide. I also wrote 5000 posts on the newsgroup about that game. The level of optimization work thrown at it was rather heavy. The primary effect of that work, in the aggregate, was to create a super-steep learning curve for newbies. Any player who followed the newsgroup regularly for 3 months could not possibly lose against almost everyone who had not - only a few particularly gifted players excepted.

Only a resolutely free-n-open culture around these things kept that fine game from being killed by this. (Incidentally, another waypoint and we-go system).

The purpose of a game system is not to raise the standard of best play to the highest possible level. That is what scripting, or similar automated "best practices" techniques, do. Nobody gains a single thing from everyone's moves in a CM game being "tweaked" to high heaven. Not a thing. Good game design had nothing to do with good CM play.

The second bad idea on the thread is one fellow expressing the sentiment that if he is "playing" the company commander, he "doesn't want" to "be" the platoon and squad commanders too. The reason military strategy games are good strategy games is because the players *do* command multiple levels.

An accurate sim of a private's actions in a civil war battle would not include a single interesting decision. For a general, it would include perhaps 2-3 decisions over the whole thing that were not immediate and automatic responses to the tactical situation.

The lives of single military officers are frankly just not interesting enough to make a strategy game out of. This is not a criticism of them. It is simply a result of the quite limited impact one person's decisions actually have on a full scale battle.

Strategy games are about a battle of wits between two opposing commanders. Their wits, the interaction of their decisions, have to decide the outcome. That is the sole attraction of the things, and it is why chess, which simulates nothing, is a fine strategy game. Meanwhile, while an accurate sim of a single Marine's life on Guadalcanal would be an excruciating ordeal, not fun.

We want a strategy game. At the moment we have one. People that want to neuter it into a mere sim do not understand what they are asking for. Good game design is a lot harder than sim design. If you get the level of aggregation just right, and the command span just right, and the degree of fog-of-war just right, you can get a good game out of it.

If you ignore those things and reduce the player to a single position, almost nothing he does will have a decisive impact on the outcome. "But I called for fire". Well, somebody else relayed the fire coordinates - maybe he screwed them up. "I told the platoon to advance down axis Alpha". But the point saw what he thought was a mine and went elsewhere. An accurate sim of a typical low-level commander's role in a WW II battle would be a mass of frustration punctuated by three randomly selected decisions actually carried into effect, one of which turned out to be wrong.

The last idea on the thread, the one that started it, I have less of a definite opinion on. I see no strong need for the kinds of additions asked for. The game would be pretty useless to me without the strategic AI. (I use it e.g. to command defenders in scenarios I design for players in campaigns). Playability is a vastly more important feature than allowing either the best possible play or the most literal implimentation of every WW II action.

If I want to represent dismounting of scouts, I put extra teams on the vehicles. I don't give a tuppenny darn whether they were "really" there. If the game effect is right - Stuart pulls up to here, men proceed on foot, men spot target, side reacts to things seen - then I do not care a whit about the rest of the details, like whether 4 men or 5 were involved and whether one of them had a scoped rifle.

On the artillery missions, they actually used lifts not continuous walks. And fine-tuning fire missions to the last shell seems ridiculous to me. Buy another module. You can have all the in-game effects, it is only competitive resource stretching in the first place.

But all of that said, if the improvements this fellow wants, or anyone else wants, would *not* hurt the AI, then fine.

[SuperSlug]

Ok so limit the scripting to designed scenarios. To give the AI a little extra smarts or to have AI controled troops react in a historical manner etc...

Personnaly I don't much care about scripting. I like the way the game works now. But I'm sick of hearing requests for new command options. It was just a suggestion to solve the problem.

However, I would still like to see the Strat AI API opened up so that other developers,like myself could write our own AI as a plugin. This of course would be in the CM3 engine rewrite time frame.

Just some wishful thinking I doubt something like that is even of the BTS radar screen.

Cheers,

[Nabla]

Originally posted by Olle Petersson:

I'm not limiting my thinking to the one minute turns, but to the response of one order.

Take for example a platoon of tanks that I want to advance along some route, seeking enemies and when found engage in combat (without advancing further).

* As it is today I'd have to plot leapfrog Hunt commands for the tanks, trying to find hull down positions or other cover when possible.

Once an enemy is encountered they either stop immediately and keep shooting 'til the enemy is gone, or they continue moving along the plotted course.

* My whish is to give an order in style of; Platoon, along this route, engage in combat, then wait for further orders.

Then the AI line up the platoon in a suitable formation and start bounding movement with overwatch.

Once enemies are encountered all tanks fire one shot, go for covered fighting positions, fire another shot, change position, ..., until the enemy is down. One section might try flanking movement and so on.

Then they wait for further instructions.

I think you get the picture...

In short: If I'm to play the role as company or battalion commander I don't want to act like a platoon or squad leader.

I see that a number of people have joined the discussion while I've been away, which is of course very nice. Let me address Olle's post first.

I see your point here: you want to be able to give even higher-order commands than what is possible in CM at the moment. We are talking about the same problem here, although our objective is different. The difficulties that arise in your scenario are exactly the same that arise when designing the strategic AI. You just want the tactical AI to take care of some strategic aspects it does not cover now.

Would you agree?

With regards to your objective: I have no strong opinion on that.

[This message has been edited by Nabla (edited 03-17-2001).]

[Nabla]

Originally posted by jasoncawley@ameritech.net:

Two bad ideas on this thread, and one on which I have less of an opinion.

Ok, let me try and address these one by one.

Scripts are a horrible idea.

I agree for the same reasons you do (as I already stated above): scripts a.k.a pre-programmed micromanagement might be needed in winning in the game, which would not be nice at all.

The second bad idea on the thread is one fellow expressing the sentiment that if he is "playing" the company commander, he "doesn't want" to "be" the platoon and squad commanders too. The reason military strategy games are good strategy games is because the players *do* command multiple levels.

...

We want a strategy game. At the moment we have one. People that want to neuter it into a mere sim do not understand what they are asking for. Good game design is a lot harder than sim design. If you get the level of aggregation just right, and the command span just right, and the degree of fog-of-war just right, you can get a good game out of it.

An interesting view, and at this point I think I agree with you on this.

The last idea on the thread, the one that started it, I have less of a definite opinion on. I see no strong need for the kinds of additions asked for. The game would be pretty useless to me without the strategic AI. (I use it e.g. to command defenders in scenarios I design for players in campaigns). Playability is a vastly more important feature than allowing either the best possible play or the most literal implimentation of every WW II action.

If I want to represent dismounting of scouts, I put extra teams on the vehicles. I don't give a tuppenny darn whether they were "really" there. If the game effect is right - Stuart pulls up to here, men proceed on foot, men spot target, side reacts to things seen - then I do not care a whit about the rest of the details, like whether 4 men or 5 were involved and whether one of them had a scoped rifle.

On the artillery missions, they actually used lifts not continuous walks. And fine-tuning fire missions to the last shell seems ridiculous to me. Buy another module. You can have all the in-game effects, it is only competitive resource stretching in the first place.

But all of that said, if the improvements this fellow wants, or anyone else wants, would *not* hurt the AI, then fine.

Ok, so these comments apply to my original idea that started the thread. First you say that "The game would be pretty useless to me without the strategic AI." This is a clear statement for the strategic AI and against my idea about a pure PBEM game. Then you examine some of my improvement ideas, and as a result you state that none of them are really that necessary in the game.

Since assessment of improvement ideas is really a matter of opinion we can freely disagree on those. However, I think that you are still dodging the main point here. Those were just examples. I did name some other improvements - trenches and complex sewer systems - in another post.

Since I know that you Jason are one of the most respected experts that write on this board let me challenge you. I'm challenging you to write down (for example ten) most important game logic improvements (that have an effect on play strategy) you'd like to see. After you've listed them, please consider how difficult it will be to make the strategic AI use these features. Ok, I know this is very difficult to assess, but make an educated guess. After that, consider which of your improvements will in fact be made.

Actually, knowing your vast knowledge base, your answer might change my own opinion on the subject.

[nijis]

BTS implemented assault boats. If the AI can't handle sewer movement or trench systems or armored car dismounts or whatever, I wouldn't think that would automatically mean excluding them for use by human players.

------------------

"I can't listen to music too often... It makes me want to say kind, stupid things, and pat the heads of people... But now you have to beat them on the head, beat them without mercy."

V. I. Lenin

[Nabla]

Originally posted by nijis:

BTS implemented assault boats. If the AI can't handle sewer movement or trench systems or armored car dismounts or whatever, I wouldn't think that would automatically mean excluding them for use by human players.

This is a very good point. Not knowing the inner logic of BTS it is impossible to say whether this particular example has been too difficult to implement or of too little value to deserve resources for implementation.

The addition of such features is in fact very much a question of customer acceptance. Let us for the moment assume that the strategic AI implementation of assault boats has been too difficult. For me it would be optimal if BTS could add features such as assault boats and leave the strategic AI implementation out if it is too hard.

However, if many such features would be added into the game, how would the players react? I mean, there would be more and more scenarios that would not be playable against AI. This would take the game towards the direction of a multiplayer game, although much more smoothly.

I admit I would very much like to hear the opinion of BTS about some of the stuff that's been discussed here. They have the real knowledge of stuff about which I can only speculate here. But I know that they are very busy at the moment with CM2 and have their priorities elsewhere. Well, on the other hand, this way I can probably speculate for a longer time.

[jasoncawley@ameritech.net]

Most of the things I'd like to see tweaked don't fit your sense of things, as held up by the AI.

I want some unit prices tweaked (wire, infantry with better weapons, flamethrowers and halftracks, uparmored tanks, cheaper fortifications - aside, and less conscipuous ones i.e. bunkers spotted about like guns not vehicles).

I want some unit types included that aren't (e.g. U.S. AA halftracks, German SiG assault guns).

I'd tweak the allowed point ratios in quick-battles (more flexibility in combined arms and mech, somewhat loosely artillery limits).

I'd increase the amount of terrain the defender has to set up in (most of the map in assaults, 2/3rds in attacks, 1/2 in probes), and make typical QB maps longer on the axis of advance rather than shallow and wide.

I'd change some of the terrain items in the map editor to allow line and tile terrain (e.g. a wall in brush) or on single side of it (e.g. field up to the edge of bocage without the "strip", or brush up to the right side of a hedge but not the left).

I'd like to see the cover model tweaked, to reduce the near-single importance of kind of terrain one is in, and increase the importance of movement state, so that e.g. running through scattered trees was almost as vunerable as in the open, while sneaking in woods would be nearly as good as being stationary in them. %exposed would decline to the stationary level with a delay-timer drop after movement ceased, or to a better level after it just slowed.

I'd slightly tweak the targeting model to avoid bunched up firing at nearest targets, especially cowering ones, when other high % exposed targets are available. AT weapons would hold fire unless they had decent hit% or were told otherwise, and squad infantry likewise at range >100m once ammo reached ~1/2 a load. Perhaps vary the "tightness" of this fire discipline / ammo use stuff with quality, greens like now but others tighter.

I'd change the groupings of skill levels to make mixtures of green and regular the most common setting, not mixtures of regulars and veterans, which too often leads to the silly all veteran version of WW II, and incidentally masks the effects of the CM command-delay model and all the nuances of the moral system. Since vets mostly do what they are told until crushed.

I'd create a "retreat" option like cease-fire and surrender now. Choosing it would prevent one from getting points from objectives but would remove the global morale penalties for exiting units. Fewer fights would end with total desctruction or surrender, more with retreat, as was historically the case. I'd let panicked to broken units be given rally-points or exit movement orders, that they would haltingly head for, instead of wallowing where they are to be butchered.

And 10th, I'd make improvements to the AI by including many rule-of-thumb cases, involving things like - artillery use (earlier, but fewer shells per target and more shifts), force mix picks (e.g. take vehicles in pairs), placement of obstacles (in belts rather than seperated, and to deny areas of cover), enemy avoidance (run from artillery fire, HQs trail others, pauses used to stay "on station" while traveling in formation, vehicles avoid places others have died as likely "kill sacks", avoid open ground more, etc), treatment of objectives (pick attack headings focused on enemy more than VLs, to clear flanks and avoid traps better).

How many of these do I expect to be done? I've seen improvements in patchs e.g. cost of fortifications. But basically I expect BTS to bring us Russians, and later DAK and Italians and such, and eventually PTO. I do expect some tweaks to items like the above list to be made for CM2, because they are sensible and do not hurt other things. Some may be too hard to program to be worth the effort, others too controversial or can-o-worms, or my opinion on them too idiosyncratic. I'll take what I can get.

I don't want anything done that will reduce playability. I am not in the least interested in "IT", aka playing out WW II with single-man counters. Monster, niche, grog-only, "Campaign for North Africa" style overdone stuff, that'd require a full time staff of a dozen trained officers to play, I want avoided above all things. Keep it simply, stupid.

That is probably compatible with a few of the things you want - say trenches. Fine by me. I do not want complexity at the expense of playability, but when added items do not reduce playability and somebody wants them, go ahead.

As for cases like assault boats, I'd rather they didn't need counters in them to move, but the AI could use them, than like now (if the "river crossing problem", aka who rows?, is the reason it can't). I'd even put up with an idealization about them, like such-n-such a side in this scenario can "move" but not "run" over water, with boats appearing when they do - if such a change made them AI-friendly.

[SenorBeef]

Originally posted by Jeff Heidman:

Funny how 95% of the psots in this thread just completely missed the point.

Jeff Heidman

Yeah, its pretty amusing.

Random Joe Poster: Wouldn't it be great if CM could implement SOPs? It would add realistic detail to the game, and make it more playable.

Random Joe Sycophant: Dude if ur gonna want that **** then u might as well ask for rail guns! go back to quake 3 and stop bothering bts!

Random Joe Poster: I also think that perhaps the infantry should have an "advance to contact" command because it seems a perfectly realistic and useful order.

Random Joe Sycophant: Yeah why don't u ask for rocket jumps while ur at it? CM is about platoons, not some little quake guy with a chainsaw! All this micromanagement is crap, go play red alert 2.

[jasoncawley@ameritech.net]

Not quite, Senorbeef. Joe poster starts with his request, and then adds - "and why not take the AI out and shoot it, to make it easier to impliment my favorite tweak?" Since lots of people use that, Joe naturally catches numerous flying brick-bats from people who rather like having an AI, and could care less about the tweak he mentioned.