Again on m.i.s.s.i.n.g. AI, some thoughts and a nice refer screen

[Kieme(ITA)]

Hi all!

I've had some good discussions about AI and what's they way it works in this new CM game. Please consider I'm not arguing about pathfinding issues, this is another history.

Actually I've some thoughts I'd like to share and a situation I've been thrown in by case that might represent the ideal description of the actual AI situation and performance.

As far as I've understood there are 2 main types of AI.

The TAC-AI works as the very base, it tells units to spot, engage and fire to the enemy, it tells to reload weapons and choose between them and ammunitions. This is all for vehicles.

For infantry it tells also units how to get inside a building (from the door) and also will give some emergency commands like "crawl there -best protected position- if fired to" and similar to units leaving KO vehicles.

The STRAT-AI is the AI planned by the scenario designer, who can work on these options:

set a command "where to go"

set a command "how much time there after/before exit from there"

set a command "how to go" (general behaviour, active-cautious-normal)

set a command "how fast to go" (speed and spotting ability balance, assault-quick etc.)

set a hide command

set an ambush command

These are the basic commands which can be used in combination to the painting instrument available to select a certain zone to go to and in which to express these basic commands.

Ok right? This is what I've understood since now.

SO, this is where the discussion might become interesting ... is that ALL? I'd like to have a reply on this foundamental question, and mine for now is: yes, that is all.

Think about this particular case (see screenshot):

First of all I'm on US side, I've created the whole level (actually a Work in Progress):

Look at it.

I've a Brad and an M1 both immobilized. I've some enemy units engaged or in sight of the US vehicles.

I've a T-72 who made its way where it's now by following the stratai orders I plotted before... it was the very last order of his chain of orders and so he will sit there (not inside a zone painted as objective to defend/keep etc.) it's just a spot and he is not engaging or seeing any enemy unit.

The minutes pass thru.... many minutes.

The T-72 will stay there 'till the end of the game I can assure you. I moved for fun 2 tanks inside its arc of fire and he killed them quite nicely (TACAI works for me).

So, what would you expect from that situation as "the best thing" or the "possible things" the T-72 could do??

My two cents would go for a nice little move the T-72 could do just in front of him... some more meters and he could get an easy kill to the brad's ass thru the collapsed wall...

Secondly T could do all the way around the collapsed building and get an easy shot on the M1 ass...

This would be the best of the best for such a move.... this could be what a human player will do.

Another kind of move could be to turn around and move to engage the M1 first, directly. Chances of success won't be high for sure, but this could be a move, a thing to do...

Finally the T-72 could go another way, searching for enemy units or something like that, just because the battle is going on somewhere else, because there are still shots being fired and people being killed...

[Keep in mind the T-72 crew is veteran and with high motivation (set by me for the scenario)...]

So much for AI existance. There's no "medium" AI that could do anything more than executing tightly what the STRATAI told to.

This is the main reason QBs don't work anymore... (first because STRATAI is absolutely necessary, and it wasn't plotted -a big and unacceptable missing for a finished game..- ; second because there's no "medium" Ai that could do even the simplest thing ever..)

In old CMs we had some very basic AI... ok it wasn't able to do any smart thing, many times it was to do the worst thing ever, but at least it DID SOMETHING.

At the present situation you will always go into something like this. AI will do the STRATAI plan, if the plan is a great one it will be nice game for sure, at least if the plan has the chance to go on until the end of the game (and this depends on how things will work toghether and is not a thing you can plan, since the case could make all the T-72 blown up in a second or they can survive and kill all M1s and then there's nothing... since even such a great acquired vantage is not to be used by AI...), otherwise you'll get to the struck situation of having the complete loss of initative by your enemy and you'll be able to do anything being sure enemies will keep there without a move.

[ August 30, 2007, 07:31 AM: Message edited by: Kieme(ITA) ]

[Kieme(ITA)]

Now I'll prevent possible objections...

-ok T-72 could not know the enemy vehicles are crippled, but anyway it should try to engage them.

-T-72 crew don't have the excuse they couldn't see the enemy directly, since a lot of fighting is going on with both M1 and Brad firing at nearby enemy units (friendly to the T-72).

And even with the rear periscopes the T-72 commander must spot the fight behind him between the M1 main gun and the building (with an inside HQ unit)...

[tarball]

If the TC shuts the engine down, he'd hear those other vehicles - otherwise, it is plausable that he could sit there and not notice those other vehicles.

Is this T-72 a part of a platoon with those other tanks at the top of the picture? If so, then he'd have radio contact with them and they would have spotted the brads and M1.

[Kieme(ITA)]

Thanks for your reply.

First those two other T-72 were KO but only after the T-72 alive arrived there, so its crew should conclude that there is for sure the Brad up there, the M1 arrived later.

Maybe the T-72 in question could not hear the M1 but cannot ignore shots and destruction being made some tens meters behind it.

Anyway tarball please note that your suggestions are referred to a possible RL situation... in this case the AI misses completely, no matter what...

[Pandur]

problem i see with that is that you have to be the "medium AI" too by giving the AI groups many little goals to get and allways telling exactly what to do. and that by time and not by acuall happenings like enemy unit crossed that line or whatever.

thats not optimal as you pointed out.

[Kieme(ITA)]

Anyone remembers that old game Operation Flashpoint?

I'm not comparing the games but for the editor, in OPF I remember you could not only plot in generale AI but you had also some more instruments which you may call "EVENTS".

You could order AI to make certain moves/decision just after some event happened.

In this case the Event might have been: "no enemy in sight for 5 minutes" and the reaction might have been: "search for enemy presence in a range of 500 meters"

Just a swift thought.

In any case I still think that some sort of basic medium AI should be added for sure!

OR give us many many more commands and options for the AI planning (although in this way you'll spend hundreds' hours to make some AI working and this is kinda frustrating...).

[KNac]

The manual says there are 3 kinds of AI, one of them being Operational AI, which is the AI you recall. I don't know to what extend this AI is programmed or tweaked, but unless the manual lies, there is such thing allready. Other thing it is broken.

[Kieme(ITA)]

Everything points out it's not working.

[birdstrike]

It's working exactly like in the CM1 engine. So it's not weaker than that.

Once a unit has reached it's destination (by AI plan in CMSF or when it reached the "flag" in CM1) it sits there waiting for kingdom come.

Some sort of "rinse and repeat" command that tells a unit to go back to step 1 or 2 or whatever in it's AI orders would be nice, I guess.

[Paper Tiger]

Nice experiment there. I've been doing some of my own too while I'm waiting for the patch. It's probably unfair to compare the AI in CMx2 to CMx1 because it's an entirely different creature.

I used to try my hand at designing scenarios for CMBB for myself and I often wished that I could issue some orders to the computer player to make it attack more effectively. With CMx2, now I can. However, the AI will execute my plan come hell or high water until the computer side is stuffed or the plan itself is finished. It plays like a robot: it will do exactly what you tell it to do and once it's finished, it will sit and do nothing unless trouble comes to them.

I'm not very sure I'd be too happy if the AI decided to act on its own initiative when it ran out of orders because, as you have pointed out, it's probably going to do something insanely stupid and get itself killed. And then we'd be complaining about that instead. "Why do my tanks reach their objective like I planned and then wander off in a random direction and getting killed!" etc etc etc...

There are real limits to what we can make the AI do at the moment. For now, my best idea is to keep the time of the scenario short so that plans can remain active. I'm not that good a player and it's hard for me to see how the situation will develop after 1 hour of gameplay. For a 20-30 min scenario, the limited number of orders we can issue can be used to make a more interesting plan. We'll just have to learn to work within the situation until the options are expanded.

[thewood]

The main problem with scripting right now is it is all time based. There are few, if any event driven scripts. Once timing is porked, the entire scenario can be written off.

[Paper Tiger]

Yes, it is. I don't think that's going to change either for some time either. Maybe we'll get it for WW2.

However, there does appear to be quite a lot of subtlty built into the AI orders which it will take scenario designers some time to master. I don't really understand them yet but the AI will execute an order in very different ways according to the instructions given. I was watching one of my experiments unfold and the AI wasn't moving on to the next objective but just pouring fire on the enemy position. When I checked the orders, I had given them lots of time to move on the next objective and a 'caution' order so before moving, they put down a lot of suppression fire. Quite nice...

The scripting for the moment is better used for the Red side because they will have less chance to "shape the battle" and react to unexpected enemy actions.

[Paper Tiger]

Replying to my own post... how narcisstic is that

While I'm waiting for the patch to get to Paradox I've been doing some more work with the Editor and the AI.

The manual says that there are three levels of AI: Strat, Operational and the Tactical and some people seem to be saying that the the operational is broken/missing. But the manual says that the operational AI co-ordinates and implements the Strat commands among the units within each group so that the player programmed Strat intentions are carried out. So, it's not broken or missing, it's doing exactly what the manual says it does.

I remember a saying "timed like a military operation" to say that a plan went well and efficiently. I think that timing based is the BEST option for WW2 as there wasn't the CC available to do anything else. I don't see any reason to believe that it's that much different today. However, with the technology that is available to modern armies today there is considerably more flexibility available, especially for the Blue team.

Without the 'issue orders while paused' in Elite (no patch yet), a half decently scripted AI can hand a RT Elite player his ass. (ie, this RT player)

[Darkmath]

I want to bump this interesting and (finally!) constructive post about AI.

First, was the test done in 1.01 or 1.02?

Do you get the same results with 1.02 , which adress self preservation AI issue?

Secondly, here is a quote from your post, Kieme :

I've a T-72 who made its way where it's now by following the stratai orders I plotted before... it was the very last order of his chain of orders and so he will sit there ( not inside a zone painted as objective to defend/keep etc. ) it's just a spot and he is not engaging or seeing any enemy unit.

I think the key of the problem may be here.

What is a objective zone for, if it wouldn't be some intructions to tell T72 to engage enemy?

A spot would just be what it is in gameplay term: a spot in which T72 won't take any initiative.The T72 have to go there , but not to defend it, so he's not required to attack Bradley.

Indeed the "medium" CMX1 AI doesn't exist, as it belongs to the AI editing right now.

[Kieme(ITA)]

Hello Dark,

thanks for your kind words.

First the test was made with 1.01. But I'm pretty sure 1.02 doesn't change the problem, since I still think this kind of problem is a core one, not easy to fix with patches.

I find this problem to be the biggest one of this game and the main reason CMSF can't express it best on single player game experience.

I wasn't clear enought about that T-72.

Actually it followed my plan and he reached the last obj zone I painted on the map.

It was quite a small zone, the T-72 decided itself where to go exactly, anyway it was just near that building, so the main move was completed OK. Sitting there it cotinued to fire upon enemies coming into its visual range. So the so called TacAI seems to work (another thing it to understand what kind of ammunitions it used against what kind of unit but this is another story).

The problem I've tryed to explain and express is the complete absence of a medium AI which could do some work within its own initiative.

My Strat AI plan told the T-72 to go there, the TacAi told it to fire upon enemies insight.

BUT the problem is that AFTER some action, after some many minutes during which the battle EVOLVED on its own, the new condition in which the T-72 found itself was completely different. And that condition is well displayed on that image.

The T-72 cannot engage the enemy from that position. It cannot fire directly to them.

The problem is that at least a sort of move would be necessary in order to engage the enemy.

AI won't do ANY KIND of movement apart what told by the StratAI.

The whole idea of the StratAI was nice, this give a lot of options for the scenaro designers although it's just it. If the battle evolves in any way there won't be any addition to what was told in StratAI plan.

That screen describes a casual situation in which the T-72 could achieve a major result, it might kill 2 more enemy vehicles, clear the Obj area and in conclusion it could reverse the sorts of the battle.

Final word: the AI does not exist. It's not enought to have the StratAI (with all it limitations) and the basic TacAI (which actually is the only way computer Ai expresses itself).

Maybe someone won't agree with comprisons with other games.

[Maybe CM is really another way of thinking about gaming but it really remains a game.

The jump to the Real Time is the definitive consacration of CM as a real member of the "wargame" or "strategic game" section of games' world. So the comparison is now much likely to happen]

I've 2 main comparisons:

First of all the "medium" AI which doesn't exist in CMSF (ok, it doesn't exist in all CM series, BUT in older CMs it did, maybe it was so simple but it was there, that's why older CMs could handle the quick battle game type which was completely broken and it still is in CMSF, now even the Quick Battles need to be scripted with StratAi by game designers!), as I said the medium AI exists in ANY OTHER wargame videogame.

It works in such titles as Total War series, Supreme Commander, Dawn of War series, just to say some examples. Units in those games won't just follow a pre-plan and fire upon enemies, they will move, try combinations of different units, they will even try differen way of attacking and will make different decisions during the battle.

This kind of AI needs a lot of work, it's OK.

CMSF was developed with a fraction of budget thatwas used for those games... I understand this, but the fact remains: AI IS MISSING. And this game still costs as much as those ones. And I'm really sad saying such things because I think that this game and kind of game as CM series could achieve outstanding results if only it could be granted more founds.

The second comparison is with the editor publicated for the game called "Operation Flashpoint".

Even if the game is a completely different kind the Editor and scripting idea are practucally the same that CMSF would like to present us.

The idea is really great, since with the right instruments some really great scenarios could be made. The problem here is that the StratAI options CMSF gives us are too few, too restricted and so scenaros won't evolve much more than some simple attack/defend situations.

In Operation Flashpoint there was a trmendous editor, even too complicated as I remember, but it was the perfect summon of instruments to make almost any kind of possile situations happen and so great scenarios were made.

FACTS:

-medium or independent AI is missing

-90% of AI performance during a battle relies on StratAI planning by the designer

-the instruments for StratAi planning are too few, too simple and in general ridiculus (and even those few are bugged!)

CONCLUSION:

-Quick Battles are no more

-Scenarios have a so narrow and thight scope due to restricted AI capabilities that the game cannot express itself.

[ August 30, 2007, 07:15 AM: Message edited by: Kieme(ITA) ]

[thewood]

The evolving battle thing is key to your comments. With only timers to set AI movement, the AI is crippled as soon as the battle moves off the time table. I have been trying to build scenarios since this thing came out and it is almost impossible to build one with any type of maneuvering over distances because you can't time them. That's on attack. AI in defence is a little easier.

I now play most of my scenarios against myself in hotseat. I don't bother with an AI anymore.

[Kieme(ITA)]

You're right, thewood.

One of my comments has been about the too weak editor options compared with the role the deisgner planning has regarding AI performance.

That's why I cited Operation Flashpoint game in which you could time as you said every single aspect of a scenario.

So that scenario designing has a so narrow spectrum of possible situations that we will soon have many almost identical scenarios, in which the map deisgn will be the only new (since I consider the map editor as one of the best deisgned things of the game)

You made another important statement: AI seems to perform better than anything in defense.

That's a direct consequence and a confirmation of what I think and expresses since now: the lack of AI makes it capable only in a defensive situation in which it doesn't need to take any initiative.

I've made a scenario in which US player has to attack Red AI defender. It worked out but it has a big limit that is depicted in the first thread of mine in this discussion...

In this scenario of mine like all scenarios being made the game will work only as far as the Ai plan is executed. From that moment on the player will play solo. This is because enemy units will just stay where they are.

This is why the tactic used by players right now is clearly: find the enemy, suppress the enemy, kill it with overwhelming superiority of firepower. This has been described by someone in another thread called 3 F or something like that. He surely made a point, but all of this is a consequence of missing AI.

The AI in a game needs personal initiative, TacAi is not enought to make a game and StratAi planned by a human is really not enought.

Even if we had an editor with hundreds of options we couldn't achieve anything better than we have now. With a really complete editor we might be able to create a great variety of scenarios but in final conclusion our achievement would be only to extend this game's life. All of this wouldn't be enought to make the game itself really better.

What would be needed is a real AI, with some programming in it. An AI that would react in some way to some possible situations.

My report on this thread is a clear example of what I am saying: there is no AI that could perform even a basic movement after the planned orders have been executed. This is unacceptable in a game since it's a major feature missing.

As the game is at this point it can express itself at its best (please note: still not "the best" but "its" best) only and only during multiplayer or "personal" hotseat.

In multiplayer all problems I cited since now desappear and will only remain mechanical (bugs) and balancing problems (which are still there).

But the fact is that this game has been sold to us with some more features such as:

-single player

This feature is missing in the extend it's not featuring any AI. This means you can play the campaign and some scenarios and even more scenarios that are being made or will be made by the community but still the AI missing problem will be there. And if you consider the limited options offered by the editor I predict we will soon see all the possible scenarios situations done and so new scenarios will be different only from the map point of view. Even force selection will be soon a standard routine since there are much fewer units types than older CMs, this means the limitations already cited will show even earlier due to this adding fact.

[ August 31, 2007, 01:02 AM: Message edited by: Kieme(ITA) ]

[Darkmath]

Not to mention that some StratAI plans are broken (Assault and Max Assault are causing the crawl of death, and should not be used as the AI hardly move from its setup zone. The others are OK).

Indeed, there is no built-in "generic" AI like in CMX1 : you have to program everything about the AI.

You have run the test to see what if the AI have not objective zone to defend.

But isn't it as relevant as a CMX1 QB/scenario without any flag, any destruction points or exit zone?

In these conditions, the "medium" AI would not help either.

Doesn't a CMX1 scenario follow a kind of minimal pre plan?

What if we run the same test with the T72 but with a painted zone at its last location?

Are we entirely sure for example, that a scenario designer cannot make an agressive CMX1 style AI (Counter attacking the flag is what I remember essentially from the CMX1 AI) with orders like Advance or such ?

However, I agree that sometimes the AI won't follow even the objective zone, as it is the case in some QBs map.

I also agree that generic AI is better than no "medium" AI at all ; even more so since the AI planning right now is time based, which restrict IMO the replayability of a scenario.

Anyway, I always play against a defending AI since CMX1, when AI sends all its squad through tall hedges to get slaughtered by a single US squad in a house (credited with 50 kills ), so the attacking AI problem...

[Behemoth]

If I may interject some comments or questions to the powers that be; Why is there not some sort of boolean order structure to the order sets, ala Steel Beasts, and/or Tacops, or at the very least a simple "stance" that dictates behavior to the AI? Hear me out, I know the limitations on dynamic AI behavior for this type of order set. Though, I do feel that the turn based play needs this and to a lesser extent the real time as well (which I feel is the only way to play this game at the moment). It would mean a lot more work for the designers of scenarios, but in the end, multi-player would shine. I'm sorry if this goes way beyond the scope for this title, I just feel sorta out of control (in a bad way) when playing this turn based, and overwhelmed when in RT. I WANT, to love this game. Sorry if this takes this thread in another direction, no hijack intended.

-Chris