AI - can anything be done?

[ackack]

Here is the situation -

Allied reinforcements must cross a long bridge.

A couple of shermans, a couple of priests.

They also have an artillery spotter.

The bridge has an antitank pillbox on the axis side.

The AI will NOT use the spotter to lay smoke on the pillbox. It just keeps trying to move tanks over the bridge which get torched every time.

Any way to make a tweak to the AI smoke usage logic??

Cheers!

[Bullethead]

ackack said:

<BLOCKQUOTE>quote:</font><HR>The AI will NOT use the spotter to lay smoke on the pillbox. It just keeps trying to move tanks over the bridge which get torched every time. Any way to make a tweak to the AI smoke usage logic??<HR></BLOCKQUOTE>

Seems to me your problem here is improper use of the AI. You have gien the AI the burden of the attack. Further, you have given it a rather complex situation, what with a bridge thrown in to complicate its "moving" thinking. Which can cloud, or even take priority over, its "shooting" thinking.

No AI, not even CM's excellent AI, does as well on the attack, especially a complicated attack, as a human. CM's does very well as AI's go, but it definitely prefers more common situations such as attacks across fields as opposed to bridges. So if you want the best AI performance, either put it on the defensive or give it relatively simple offensive tasks.

That said, on the subject of the AI's smoking habits, I don't see there being a problem with under-use--rather, the reverse. I find the AI firing or popping smoke all over whether I ask for it or not, greatly hindering the fire of other friendly units.

As to it using massive arty smokescreens, it does that as well. But there are 2 factors to consider. First, the AI knows that arty's most potentially decisive use is killing things, and that each smoke round fired reduces the number of available HE for this primary mission. Thus, it seems the AI much prefers to use mortars for smokescreens, with which I concur.

Second, there is the time required to get an arty smokescreen in place. In most cases, you must await arty smokescreens for 3-4 turns. What if the AI doesn't think it has the time to waste?

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

-Bullethead

Visit the Raider Operations message board at www.delphi.com/raiderops

[Maximus]

Actually I have been thinking about this lately too. I was basically wondering if BTS could make any further tweaks to the Operational and Strategic AI. The Tac-AI is pretty much on target now with the v1.05 patch. However, I *still* do see some eratic behavior at times. Namely the "breaking" of infantry units under fire when they haven't even taken a casualty.

If there was a way to improve the Strategic-AI and Operational-AI to offer an increased effort in a computer attack then that would be most appreciated.

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

All that sleeps, awakens...

All that awakens, hungers...

All that hungers, feeds...

[McAuliffe]

<BLOCKQUOTE>quote:</font><HR>Originally posted by ackack:

(...) The AI will NOT use the spotter to lay smoke on the pillbox. It just keeps trying to move tanks over the bridge which get torched every time.

Any way to make a tweak to the AI smoke usage logic??

Cheers!

<HR></BLOCKQUOTE>

I had this problem too while testing a scenario. Then I gave the AI a mortar spotter with lot of ammo and suddenly the valley was full of smoke, Seems AI doesn't want to spent precious high caliber ammo at smoke missions, so try low caliber and much ammo.

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

http://users.pandora.be/aneric/index8.htm#Projects

[Dr. Brian]

Sadly, the game engine for CM does not use AI or more appropriately, an "expert system." I surmise, from what you tell me, and what I've seen, that it uses a series of code that acts and reacts to certain actions.

To get some logic going, it needs to use soemthing like Bayesian Networks (based on Bayes Theory) or actual Fuzzy Logic (developed by Zahdai <sp>).

Just my 0.02

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

Doc

God Bless Chesty Puller, Wherever He Is!

[Schugger]

<BLOCKQUOTE>quote:</font><HR>The AI will NOT use the spotter to lay smoke on the pillbox.

<HR></BLOCKQUOTE>

What kind of artillery spotter is this?

IIRC not all kinds of atillery can fire smoke rounds.

[Abbott]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Dr. Brian:

Sadly, the game engine for CM does not use AI or more appropriately, an "expert system." I surmise, from what you tell me, and what I've seen, that it uses a series of code that acts and reacts to certain actions.

To get some logic going, it needs to use soemthing like Bayesian Networks (based on Bayes Theory) or actual Fuzzy Logic (developed by Zahdai <sp>).

Just my 0.02

<HR></BLOCKQUOTE>

Jeez Doc, write one durn paper on "Expert Systems" and off you go wanting it used on everything!

[This message has been edited by Abbott (edited 09-14-2000).]

[Ben Galanti]

Dr. Brian,

What do you mean when you say:

I surmise, from what you tell me, and what I've seen, that it uses a series of code that acts and reacts to certain actions."

Do you mean that CM is preprogrammed to take certain actions when it sees certain situations? An 'If A then B' kind of thing? I very well may be misunderstanding this quote though. If that is what you meant, I'd be curious as to why you think this. I have no idea as to the inner workings of CMs AI (nether does anyone else on the board except for Charles and Steve) but the AIs behavior in CM hasn't seemed this way to me. Just wondering what you see coming from an academic background in AI.

[howardb]

I've just made a scenario with that problem and you just can't expect the AI to defeat a river crossing + a pillbox. There's two things to remedy this:

A) Change the pillbox w/Pak gun.

Give the AI plenty of units.

The problem with answer B is that the scenario get's unplayable in anything but vs cpu.

[Pham911]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Dr. Brian:

Sadly, the game engine for CM does not use AI or more appropriately, an "expert system." I surmise, from what you tell me, and what I've seen, that it uses a series of code that acts and reacts to certain actions.

To get some logic going, it needs to use soemthing like Bayesian Networks (based on Bayes Theory) or actual Fuzzy Logic (developed by Zahdai <sp>).

Just my 0.02

<HR></BLOCKQUOTE>

I'm happy to announce that Steve and Charles have given me authorization to offer you a place at BTS.

Your official job title will be Head Surmiser, and you will be running the B and S divisions of BTS.

Normally, this kind of job is reserved for people who know how BTS coded its product, but given your vast body of computer game AI work, and what a great guesser you've shown yourself to be they felt it was appropriate to offer you this position without the customary interview.

They look forward to seing you around the office, and ask that you bring donuts to mondays meeting.

Regards,

Pham911

Job Placement Services, Inc.

[Dr. Brian]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Ben Galanti:

An 'If A then B' kind of thing? I very well may be misunderstanding this quote though. If that is what you meant, I'd be curious as to why you think this. I have no idea as to the inner workings of CMs AI (nether does anyone else on the board except for Charles and Steve) but the AIs behavior in CM hasn't seemed this way to me. Just wondering what you see coming from an academic background in AI.<HR></BLOCKQUOTE>

Ben,

I posted a while back on this subject. MajorH and BTS staff responded with some great input onto the workings of CM (hey Pham 911, blow me, after your grow up). We talked about AI, Expert Systems, and such. To paraphrase, it's not a true AI that uses fuzzy logic (as defined by Zadeh (1978)). That leaves only one likely alternative. A series of codes that acts, based on the situation. IT was discussed, that if in fact the computer opponent reacted and 'learned" from it's own experiences, it could in fact be considered a form of expert system. However, this is more proprietary to BTS (which Pham911 is now a hot shot at, or at least, the official BTS viper tongue).

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

Doc

God Bless Chesty Puller, Wherever He Is!

[howardb]

Now now kids calm down...

[Wayne]

I am a total illiterate when it comes to computers, or anything that uses logic, but I've heard of simple computers that solve problems over time. Is it possible for CM's AI to "learn" how to play a better game of CM over time and many plays? Does it learn what not to do or do in certain situations, unlike myself?

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

Blessed be the Lord my strength who teaches my hands to war and my fingers to fight.

[L.Tankersley]

No, CM's computer player does not get better with time. It is possible to construct systems that can learn by experience (e.g. trainable expert systems) but that's not how CM works. Every scenario (and every playing of a given scenario) is a completely new experience for CM; it has no memory of what has gone before. [it may or may not have a memory of what has gone before during a scenario; I don't recall ever seeing Steve or Charles comment on this one way or another.]

Brian is being a bit pedantic when he says CM does not use AI; while this may be true under a strict academic definition of "artificial intelligence," within the domain of computer games the term "AI" is commonly used to refer to any of a number of mechanisms for generating command inputs for the computer opponent, including for example scripts, finite state machines, neural networks, and rule-based expert systems.

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

Leland J. Tankersley

[Schrullenhaft]

Just to quote on the subject of the AI:

I find it interesting that no one (to my knowledge) in the gaming industry has yet to come up with an AI that learns and adjusts itself to the player (as we would "like"). I assume that there are several reasons for this (all of which are my highly subjective opinion):

1) The amount of time to code up such a computer opponent is beyond the "developement window" where a company could financially gain from the cost of such a long-term project.

2) Most games on the market probably wouldn't really benefit from such a computer opponent. Most consumers probably wouldn't appreciate, or more importantly pay for, the amount of development time to create what may be a marginally better opponent. BTS has stated in the past that it would take them 6 months to make a marginally better AI - and of course there would still be complaints.

3) Head-to-Head play has taken over the market and computer opponents are becoming less of a factor in a game's success.

4) "Expert systems" are usually employed in a limited number of situations where A LOT of data can be gathered or input into the system to allow it to quickly make adjustments and decisions that are often too menial or that a human couldn't perform fast enough. Apparently it takes years to develop such systems and often once they are employed it may take them quite awhile before they are accomplishing their designed task to the level that was intended. Despite the amount of detailed statistics and paperwork maintained during and after WW2, the information gathered does not make for "expert systems" input at a tactical combat level (maybe at a "logistical level"...).

5) Current computers probably couldn't keep up with the number of calculations and data keeping/referencing that would be necessary for an expert system. Remember that CM also does not use a perfect ballistics/trajectory/LOS model because it would be impossible to play on anything as slow as a high-end consumer PC.

In my opinion "Dr. Brian" is arguing semantics when it comes to "AI" and "fuzzy logic". Yes, he may be literally correct in his insistence that CM doesn't qualify in its use of "fuzzy logic". But this industry is subjected to hyperbole of one sort or another and BTS hasn't been completely exempt from this - they have to market a game after all. I don't fault them for this since they have created the best tactical computer opponent available.

Expert systems just aren't going to see the light of day in a $45.00 wargame (at least not in the near future). Now if someone can get BTS a DoD contract worth $50 Million (and the rights to use it commercially)...

OK... I'm done with my belated (and belabored) ranting and acting like a blowhard.

[This message has been edited by Schrullenhaft (edited 09-15-2000).]

[Pham911]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Dr. Brian:

(hey Pham 911, blow me, after your grow up.<HR></BLOCKQUOTE>

Blow me *and* grow up in the same sentence? Tsk. Tsk. Tsk.

You could use some of that AI stuff yourself.

[Mannheim Tanker]

I'm working on a project right now that is using an expert system (I can't say what it's doing, or I'd have to kill you ) but I CAN tell you that expert systems and "learning AI's" can REALLY hit a CPU hard. In fact, we've designated a server just for the expert system. Aside from minor tweaks, I'm guessing that BTS has optimized the AI so that we get a relatively smart opponent that doesn't require a Cray to run on. Maybe in a few years we can expect an expert system-based AI. With most commercial expert systems running $2000 (eg Eclipse) -$100,000 (eg Art*Enterprise) per license, I wouldn't expect it too soon, however

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

...Damn the Kubelwagen from hell!! - Von Brizee

[Wayne]

Damn! Then could someone tell me why my game's AI keeps beating me, pursuing my men with a maniacal vengence, following my tanks with destructive barrages of artillery and hunting my leaders with determination that rivals the best human player on this board. Or am I stupid?

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

Blessed be the Lord my strength who teaches my hands to war and my fingers to fight.

[Strat]

The neural network is the only fuzzy logic I have worked with. It involves taking sample data cases, thousands, of known inputs with a corresponding output or outputs. Each case contains a number of variables, (nodes), defining an input and one or more variables, (nodes), defining an output. The input variables are wieghted across all of the individual cases, non-linearly or using multiple layers. This produces a generalized or fuzzy output. If the network is large, with each data set containing 50 or more variables (nodes) and thousands of samples cases, it takes weeks or months on a fast computer to learn the model. Real time learning is not practical using the neural network technique but it appears to be the best approach to A/I. After learning the model the calculations are very fast. It depends on the resolution though. Very high resolution A/I would be impractical with todays technology. Maybe when the terra hertz machines come out one day, maybe. CM's A/I is ok. When attacking it does tend to charge the flag, taking the shortest route. When defending it can be overrun, easily. It appears to react either slowly or caustiously in most situations.

Strat

[Dr. Brian]

<BLOCKQUOTE>quote:</font><HR>Originally posted by L.Tankersley:

Brian is being a bit pedantic when he says CM does not use AI; while this may be true under a strict academic definition of "artificial intelligence," within the domain of computer games the term "AI" is commonly used to refer to any of a number of mechanisms for generating command inputs for the computer opponent, including for example scripts, finite state machines, neural networks, and rule-based expert systems.

<HR></BLOCKQUOTE>

Although I don't think I'm being "pedantic" with my knowledge, I do, 110% agree that the term "AI" used in computer gaming is so far off from the definition in research and academia.

"AI" in the mainstream computer world is thrown around very flippantly.

But, for the most part (except being pedantic) I agree with Leyland, although I do disagree that any rule-based expert systems and nueral nets are being used, to my knowledge. If anyone can tell me of some games that are using rule-based expert systems or neural nets, I'd like to know. The game engine would be interesting to study.

However, I am hopeful, that in time, they will. Many indivuduals are becoming better at these systems, and will be bringing that knowledge and experience (i.e., the academic definition of AI) to the gaming industry in a few years.

So, the future of gaming looks really good.

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

Doc

God Bless Chesty Puller, Wherever He Is!

[Dr. Brian]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Pham911:

Blow me *and* grow up in the same sentence? Tsk. Tsk. Tsk.

You could use some of that AI stuff yourself.<HR></BLOCKQUOTE>

You know, your absolutely right.

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

Doc

God Bless Chesty Puller, Wherever He Is!

[This message has been edited by Dr. Brian (edited 09-16-2000).]

[Dr. Brian]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Schrullenhaft:

5) Current computers probably couldn't keep up with the number of calculations and data keeping/referencing that would be necessary for an expert system. Remember that CM also does not use a perfect ballistics/trajectory/LOS model because it would be impossible to play on anything as slow as a high-end consumer PC.

<HR></BLOCKQUOTE>

Schrullenhaft,

Man, you hit the nail on the head there in your entire post (I just snipped that one quote though). As Leyland pointed out, and I am aware of, the "AI" definition used is different. But you're on the money. They need to market it!

Regarding the quote above, the beauty of rule based systems like an expert system, is that it DOESN'T need to go through tons of data (That's where that Bayes Theory comes in). Well, at least some expert systems.

So what it comes down to, and you clearly spelled it out, is the costs! There just isn't enough "experince" out there to do it cheaply.

But, in time, as AI, neural nets, expert systems, and the like are common courses in colleges around the world, that expereince will be "cheaper."

Right on with all your comments though!!!

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

Doc

God Bless Chesty Puller, Wherever He Is!

[This message has been edited by Dr. Brian (edited 09-16-2000).]

[Dr. Brian]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Mannheim Tanker:

I'm working on a project right now that is using an expert system (I can't say what it's doing, or I'd have to kill you ) but I CAN tell you that expert systems and "learning AI's" can REALLY hit a CPU hard. I

<HR></BLOCKQUOTE>

Dang! THis sounds like a monster...Can you give us a hint???? Please.

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

Doc

God Bless Chesty Puller, Wherever He Is!

[L.Tankersley]

<BLOCKQUOTE>quote:</font><HR>Originally posted by Dr. Brian:

Although I don't think I'm being "pedantic" with my knowledge, I do, 110% agree that the term "AI" used in computer gaming is so far off from the definition in research and academia.<HR></BLOCKQUOTE>

Embrace the pedanticism! Where do you keep your extra 10%?

<BLOCKQUOTE>quote:</font><HR>But, for the most part (except being pedantic) I agree with Leyland [Leland -- LJT], although I do disagree that any rule-based expert systems and nueral nets are being used, to my knowledge. If anyone can tell me of some games that are using rule-based expert systems or neural nets, I'd like to know. The game engine would be interesting to study.

<HR></BLOCKQUOTE>

IIRC, Bungie/Take Two's game Oni (now in beta) was advertised as using "constrained neural nets" for the opponent "AI" (if you'll forgive the term). I don't know if they stuck with it, or whether it's really a fair description.

A few years back I got a computer Bridge game that was supposed to be able to learn and improve over time. It sucked so badly at the start, though, I never played it enough to see if that was true or not. I think it used neural nets, but I may be misremembering.

Eight or nine years ago I was lead developer for a Monte Carlo timestep multiwarfare simulation that used a sort of bastardized expert system for all tactical decisionmaking. We had NASA's CLIPS as the inference engine, operating on four "meta-rules" that in turn operated on a wide set of "rules" adapted from various tactical publications and other sources. There was also a multi-level state transition machine for every entity used to track states, goals and subgoals. (Think of it as the "kitchen-sink" approach.) Back then we were running (and occasionally thrashing) on Solaris and HP/UX workstations, but I'm sure it would run just fine on today's PCs given the increased storage and computing power now available. So I think it's at least technologically possible for games today to take advantage of more advanced techniques; the problem is that in most game development the computer opponent doesn't get dealt with until very late in the process, and grandiose plans often end up tossed in the wastebin as deadlines approach in favor of safer, easier methods.

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

Leland J. Tankersley

[Mannheim Tanker]

Sorry, Dr. Brian, but I can't say too much for fear of losing my job! I can say that it's in the area of law enforcement, but that's about it.

An interesting side note, is that one of the most expensive expert systemsm, Brightware's Art*Enterprise (@ $100,000 a pop) is used almost exclusively by the financial industry. Most banks use this (or a similar system) to determine whether or not you get a bank loan. So be nice to HAL if you want to get that dream home