What wargame has the best AI?

[Hertston]

I think people may be missing the significance of Deride's "of course, my algorithm must be allowed to run as long as it wants whenever making any decision", particularly in the context of "work on a wide group of machines within an acceptable performance watermark".  There were chess algorithms that could beat any (to all intents and purposes) player twenty years ago, given unlimited time.  Positions were actually assessed by professional players and their teams just by setting them up and letting whatever analysis software they used run for a week.  But in an actual chess game you don't get unlimited time.  Ditto I suspect with turn-based wargames.. but who would use such an algorithm as an extra setting if each AI turn takes 12 hours or more to run... [;)]

[ravinhood]

Well people play PBEM games of WitP for a year...why not 12 hour turns for a AI?
 
@jvg I still accept your apology although you wish to refrain from the embarrassment of&nbsp; saying it in public thanks for the PM. <snicker>

[Hertston]

ORIGINAL: ravinhood

Well people play PBEM games of WitP for a year...why not 12 hour turns for a AI?

Because extra sales to those who bought a game that they would not otherwise have purchased specifically to play against such AI wouldn't cover the price of a couple of pizzas, let alone its development.

[ravinhood]

How can you make that ASSUMPTION without proof? I can counter by saying you're wrong and sales of a game like that with an AI like that would skyrocket.

[Jeffrey H.]

I am also very curious how you were able to "jump" from game AI to the "real" world of business applications of AI. How did you score that ? Friends and associates, word of mouth, papers published ? How ?

[ravinhood]

Same way Herston did, by assumption of course.

[sterckxe]

ORIGINAL: Deride

ORIGINAL: sterckxe
1) Neural networks are rather good at single tasks, answering a simple yes/no question
2) Neural networks need to be trained/seeded with lots of accurate input data and output result data in order to be any good
3) Neural networks only work fine with a limited amount of input factors.

I don't agree with this. Back-propagation neural networks can provide an unlimited number of inputs and an unlimited number of outputs. For example, my thesis work was done on a neural network that took printed articles (such as the Wall Street Journal) and parsed out what part of speech each item was. (For example, if you gave the network 'the big cat', it would return 'defintive article, adjective, noun.') The outputs ranged from 1-48 different parts of speech (for the life of me, I can never remember them all, but they include 'proper noun', 'noun', 'gerund','transitive verb','intransitive verb', etc. etc.)

Ok, nice to know that the practical limitations which existed over a decade ago - and on which I based my "rules" - aren't a problem anymore today. We never looked at it from the theoretical perspective of what NN's could do, as a business we were only interested in the practical applicability.

Greetz,

Eddy Sterckx

[Deride]

A few points:
&nbsp;
1) As for the TS:JC AI. I never said it was the best, I just had mentioned that it used genetic algorithms. A reviewer calling it predictable is probably more talking about the setup -- which was fixed not random. So, each level was essentially predictable. Though we did this to simulate&nbsp;some actual battles, and we thought it was a lot of fun and interesting.
&nbsp;
2) Herston hit the nail on the head. If you let my algorithm run as long as it wants, I can always run a simple 'brute force' algorithm that allows me to fully investigate every move, every counter move, etc. etc. In TS:JC, unfortunately, this would take about 78 trillion years to run on each turn -- still probably not appropriate for PBEM &nbsp; Now, I'm not arguing that I should be allowed to run a brute force algorithm for everything, but if you give me several hours on each turn with a predictable piece of hardware, I'm gonna make an AI nearly impossible to beat.
&nbsp;
3) Ravinhood, I understand where you are coming from. And, thus, the whole purpose of my response. AI is *hard* to nail not because developers are trying to make an AI that is a hard opponent to beat. AI is *hard* because you are trying to make an opponent that *you can beat* while giving you a great challenge and a fun experience. Through internal testing, alpha testing and beta testing, you have to constantly tweak the AI against various opponents to get this perfect balance -- and it is extremely hard.
&nbsp;
Deride

[ravinhood]

@Deride what I've always asked though is that optional EXTREMELY HARD setting that is beyond the normal Hard and Hardest modes. That EXTREME setting soto speak or extended sliders that can increase the power of the AI by ability to play as well as the usual handicaps and tweaks the regular hard and hardest modes give. Using Galactic Civilation as an example it has 13 difficulty level settings.....why can't wargames?

For me nothing burns me more than when I see some newbie yoyo complain that an AI is too HARD. I cringe when I read those and I cringe even more when I read a developer decide to dumb down the AI for the sake of that one newbie complainer. This is what Slitherine/Iain did to Spartan when they did patch 1.017 If anything AI's need to remain or be built harder at the higher levels of difficulty. The newbies aren't going to play these or shouldn't be playing these anyway and the very highest level should be an extreme that is very hard to beat.

Warlords IV at it's highest difficulty is one of the toughest challenges I've seen in an AI with all AI opponents in the game (not just 1 against 1 like some of these players who claim they can beat the hardest difficulty play against). I play against the farm on the highest difficulties that the game(s) allow. To me beating that is an accomplishment and something I can do with the majority of games I play. RTW was a walkover, M2TW was a better challenge, but, still no less I found the tactic to get an upper hand and thus gain ground an eventually beat it at the highest difficulty the very first game. Though it did take two games to beat it as Hungary, tougher starting position and more spread out city/castle distances. Hungary you have to play more as a grunt rush than strategic attrition.

No game should be made that someone like me can beat the highest difficulty the very first game though I don't think. I also would rather not just have to rely on the AI getting more resources and greater advantages than myself. I'd much rather rely on a smart AI as you say you can build. The one thing I'm not though is a great chess player else I would want to play against Big Blue in chess. I guess what I am looking for is something near Big Blue in a wargaming experience. Playing solo vs an AI should be entertaining as well as a challenge. The challenge is what is missing is many AI's. I see them more as deterents to the end than a competitor. I want to see AI's play to WIN not just delay me to the inevitable

[Jeffrey H.]

ORIGINAL: ravinhood

Same way Herston did, by assumption of course.

Misdirected. I was asking, (trying to ask) Deride....sorry about that.

[ravinhood]

No need to be sorry

[Deride]

ORIGINAL: Jeffrey H.
I am also very curious how you were able to "jump" from game AI to the "real" world of business applications of AI. How did you score that ? Friends and associates, word of mouth, papers published ? How ?

Jeff,

Actually, I started on the business side and then headed towards the video game side. I've been in consulting for most of my professional career, and I have a lot of contacts at many major companies around the U.S. Fortunately, I've been able to provide strong enough results that I often get called in for particularly difficult situations to provide input/advice. In the particular case of using TS:JC'S AI, I was working with the client to do a vendo selection, and we actually started off recommending iLog -- but they flinched at the price. So, we proposed doing it in-house.

Deride

[Deride]

ORIGINAL: ravinhood
@Deride what I've always asked though is that optional EXTREMELY HARD setting that is beyond the normal Hard and Hardest modes. That EXTREME setting soto speak or extended sliders that can increase the power of the AI by ability to play as well as the usual handicaps and tweaks the regular hard and hardest modes give. Using Galactic Civilation as an example it has 13 difficulty level settings.....why can't wargames?

ravindhood, I understand where you are coming from, but you have to understand that AI in games (at the current time at least) is all about taking short-cuts.

Let's take TS:JC as an example. If I have 30 units each that can make up to 30 different moves, I have 30^30 combinations of moves that I can make. That's 2.05 x 10 ^ 44 -- a really big number. If I'm able to process, say, 100 million of those moves every second (much faster than any PC I know about), it will take me 2.05 x 10 ^ 36 seconds to process the move. That's 65,005,073,566,717,402,333,840,690,005 years.

Just stop and think about that for a second. For me to just process all of my potential moves and pick the best one, I'm talking about an amount of time that would lay all of us to waste for millions of generations. So, as an AI developer, what do you do? Well, again, you take shortcuts. You develop algorithms that look at only a very small fraction of those moves.

So, again with the TS:JC example, we used a genetic algorithm that is able to process in about 5 seconds -- and we look at approximatley 10,000-20,000 combinations. Now, if you were to allow this AI to spend time looking at 20,000 - 40,000 combinations ... or even 1,000,000,000 combinations, you are still well under any decent amount of combinations. So, letting the AI run much longer will result in little, if any, improvement of the results. For TS:JC, the difficulty setting just allowed the algorithm to run a bit deeper and with less randomness in coming to the 'best' move. But, even at the hardest potential setting, the AI would do not better than aproximating the 'best' move.

Of course, the other approach to making the AI harder to beat is to give it more units or other advantages. As a purist, I really don't like this because, to me, it really is cheating.

Deride

[ravinhood]

So because you are a purist you can't make&nbsp;at least&nbsp;ONE cheat mode of play in your games. Because I'm one of those that doesn't mind the "handicaps an advantages". I never really saw it as cheating like some others do. I just look at it as another level of play against a computer. I'd rather play against a high level cheat than to complain how lousy the game is because it's not challenging enough. I also can understand what you are saying about shortcuts lord knows I wouldn't want to wait that many days to take a turn. But, back in the C64 days vs SSI's civil war games we did wait several minutes sometimes to take our turns again. I think too many live in the McDonald age of computer gaming as well. Too much of this fast paced gaming, wanting an AI to make its move in seconds. I certainly can stand a few minutes wait for the AI to take it's turn if it's at least a decent Ai while it does this. Just some suggestions from an old grog.

[JudgeDredd]

But what he's just pointed out is exactly that the few second/minutes wait makes very little progress in the AI being able to make better estimates....the numbers are just to big to crunch.

And in the C64 days, you were probably looking at a coupel hundred calculations, not 10s of thousands as now.

Surprisingly, I agree and also don't regard allocating extra units/resources as cheating. I also think it's just another level of game. Just so long as I know that's what is happening.

[Deride]

At the end of the day, I think most developers will go ahead and add in some level of 'help' for the AI. Extra units, more resources, 'scripted events' (such as new units appearing out of no-where), etc. And, for now, that is what we need to do.
&nbsp;
Just as an aside, for anyone that is interested in AI at all, I would stronlgy suggest the book Singularity by Ray Kurzweil. He basically plots out how advanced computer AI has come and talks about the future of AI. He predicts the time when a single computer will have the same 'raw horsepower' as a single person, and when a single computer will have the same 'raw horsepower' as all&nbsp;people on Earth combined -- surprisingly within the next 50 years or&nbsp;so. Many call him an optimist, but even if you give him 1,000 years for it to happen, it does speak to some rather fascinating things.
&nbsp;
Deride
&nbsp;
P.S. I hope no-one is taking any of my comments as argumentative or anything else. Just giving my personal perspective on wargaming AI. This is certainly an area that I really enjoy -- especially since it is fun to talk to my academic friends who can beat me up when it comes to theory, but I can beat them up when it comes to actually doing it

[Jeffrey H.]

ORIGINAL: Deride

Just as an aside, for anyone that is interested in AI at all, I would stronlgy suggest the book Singularity by Ray Kurzweil. He basically plots out how advanced computer AI has come and talks about the future of AI. He predicts the time when a single computer will have the same 'raw horsepower' as a single person, and when a single computer will have the same 'raw horsepower' as all people on Earth combined -- surprisingly within the next 50 years or so. Many call him an optimist, but even if you give him 1,000 years for it to happen, it does speak to some rather fascinating things.

Yeah, that was my next Q, a recommended book on the subject.

BTW cheating AI is something that does really get under my skin. Additional force balance is ok, but boosted capabilities of like units and most importantly, intel that I don't get really bothers me.

When the computers "army", (or whatever unit) somehow magically takes down your own equivalent unit at unusually good ratios, then the BS meter starts going off.

But when the AI sees everything and you get FOW, oooohh that stings !

[Banquet]

ORIGINAL: Deride
Let's take TS:JC as an example. If I have 30 units each that can make up to 30 different moves, I have 30^30 combinations of moves that I can make. That's 2.05 x 10 ^ 44 -- a really big number. If I'm able to process, say, 100 million of those moves every second (much faster than any PC I know about), it will take me 2.05 x 10 ^ 36 seconds to process the move. That's 65,005,073,566,717,402,333,840,690,005 years.

Even if/when computers have the sort of power to number crunch like that I don't think it's the answer to making a good AI. In the above example the AI wouldn't be planning ahead, but would be making each move or turn in complete isolation. Even if future turns could be factored in it would, if it worked perfectly, be like playing a robot who made no mistakes. I don't think it would be fun to play against.

Human players don't analyse turns to such an extent. In my opinion what current AI's lack.. and what would make them not only better, but more fun to play, would be the ability to look ahead and make it's turns based on the overall objective. I.e, it might create a feint attack, or hold back to gather reinforcements. Another aspect would be learning from the humans play style. Some people might be able to beat an AI straight away but many (me included) start to beat an AI once we see how it plays. Generally the AI won't vary it's style too much from game to game. Humans learn from that. An AI needs to be able to learn the way a human plays too, and adjust accordingly. It would feel more like playing a good human, and less like playing an automaton.

Of course programming that would no doubt be incredibly difficult. I don't doubt it would take quite a powerful computer to do it, but I don't think it would require the sort of grunt that calculating 2.05 x 10 ^ 44 different random moves would.

[Deride]

ORIGINAL: Jeffrey H.
Yeah, that was my next Q, a recommended book on the subject.

The Kurzweil book is a good read for anyone -- no need to be a developer or highly technical person.

For someone interested in learning more about AI in games, I highly recommend the 'AI Game Programming Wisdom' series edited by Steve Rabin. This is all a lot of academic stuff, so you need to have a decent background in algorithms, data structures and general AI approaches to actually implement anything.

Also, if you have anything specific that you would like to learn about, please let me know -- I can point you in the right direction (hopefully). I'm also happy to hear about any ideas, etc. I always love a problem to think about

Deride

[HansBolter]

ORIGINAL: JudgeDredd

But what he's just pointed out is exactly that the few second/minutes wait makes very little progress in the AI being able to make better estimates....the numbers are just to big to crunch.

And in the C64 days, you were probably looking at a coupel hundred calculations, not 10s of thousands as now.

Surprisingly, I agree and also don't regard allocating extra units/resources as cheating. I also think it's just another level of game. Just so long as I know that's what is happening.

I have never decried it as "cheating". I simply detest it being used to an "nth" degree where it quickly reaches a level of absurdity.

HOI2 is one of my favorite examples. I wouldn't mind playing a German against an AI Soviet that gets 500 divisions.

I mind greatly playing a game that has the audacity to depict itself as WW2 based that gives the Soviets 1000 divisions in order to allow it to be competitive.

I will never mind giving the AI a few extra resources SO LONG as those resources represent something PLAUSIBLY REALISTIC.

As soon as the gift to the AI reaches the level of COMPLETELY UNREALISTIC the game is RUINED.