What wargame has the best AI?

[tomcat666]

What game according to your experience has the best AI?, and what matrix game has the best?

A question for the designers:

And when will designers take advantage of the new techniques in programming such as heuristic algorithms, adaptive evolutionary learning and fuzzy logic ( used in Combat Mission)?

The processing power of PC’s has now made such techniques practical.

[Plodder]

Conquest of the Aegean in both categories..

[Gil R.]

A few months ago there was a poll on this topic at Wargamer which was deeply flawed in several respects, especially because it mixed strategic and tactical games. As a developer, I can tell you that comparing those AI's is comparing apples and oranges. Also, I can tell you that sometimes players -- and, in one case I can recall, reviewers -- don't realize which elements of a game are run by the AI and which aren't. (This is sometimes known only to the programmer.) The better question, then, might be about which games provide the greatest challenge as one-player games and can be replayed the most times without becoming too predictable.

[sterckxe]

ORIGINAL: Gil R.
The better question, then, might be about which games provide the greatest challenge as one-player games and can be replayed the most times without becoming too predictable.

Agreed, but the answer remains the same : Conquest of the Aegean [;)]

Greetz,

Eddy Sterckx

[Sarge]

Conquest of the Aegean

[HansBolter]

Actually the best hasn't been released yet.

It's Battles from the Bulge only because it includes improvements on Conquest of the Aegean.

[Deride]

I had responded to the other AI-based thread that you posted as well. But, I think your question isn't really symetrical. Using the most 'modern' AI techniques is certainly not always going to produce the best AI.
 
For example, the standard pathfinding algorithm (A*) will often result in the optimum path to get from point A to point B, but players often will disagree with the path. Another example I know is in the creation of random numbers. 'Truly' random numbers will still have streaks in them -- e.g., 1, 0, 1, ,1 ,1 ,1, 1, 1, 0, 1 ,1, 0. That string of 1's is entirely random, but players will see this and call the AI 'garbage.'
 
I've worked with a number of AI algorithms and approaches (my Master's Degree was in the use of recurrent neural networks to solve human language problems.) In the end, the 'best AI', has to achieve goals that are beyond simply being 'smart'. They have to:
 
1) Meet the expectations of the user (for example, AI's that are too smart are often acused of cheating)
2) Adapt to the playstyle of many users (to be both a challenge and make the game possible to beat)
3) Be just difficult enough to put up a challenge -- but not hard enough to be frustrating (this is a *game* afterall)
4) Work on a wide group of machines within an acceptable performence watermark (which is HARD to accomplish with such variety of memory, processing power, etc.)
 
Trust me, I can always build an AI that can beat you at just about any game (Of course, my algorithm must be allowed to run as long as it wants whenver making any decision.) But, balancing the above 4 is what makes a truly great AI vs. a hard-to-beat AI.
 
Deride

[sterckxe]

ORIGINAL: Deride
For example, the standard pathfinding algorithm (A*) will often result in the optimum path to get from point A to point B, but players often will disagree with the path.

That's why Zeus invented waypoints [;)]

Seriously : I posted this somewhere else a couple of weeks ago and thought that with your academic background you might want to comment on it - or straighten me out alltogether [;)]

-

Every decade or so there's a neural network hype - at that point every software developer starts putting "neural network" on the item list on the box. "Neural network" in a business program usually meant : has an ordered list of most recently opened files by the user or the wizard remembered the options you picked last time or the program remembers the user-settings. Not unimportant from a UI pov, but this has nothing to do with neural networks. I took a good hard look at *real* neural networks for a serious banking business project back in the nineties and learned three important things :

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.

If you want to use neural networks for more general purpose problems or don't have accurate historical input AND output data or have a zillion factors to take into account for a single answer : forget it

Example of what it can do - and does amazingly well. Back then I worked for a company which had a cafeteria in which each day they would serve 3 different lunches so I seeded the neural network with the meal info broken down into categories (meat/fish, vegetable, side-dish stuff, dessert, ...) and which meal I picked. After inputting months and months of data the neural network managed to predict with uncanny accuracy which meal I was going to eat that day based on the
ingredients of the 3 on offer. You didn't have to program anything for this, nor did you have to define the logical path and links between the nodes, you just had to provide input and output data.

More practical business applications are in the actuary, mortgage, loan & credit, insurance and other situations where lots of factors come into play, but you only need a single yes/no answer in the end. What you do with a neural network is that instead of heavy math and guesswork to determine the weight of each factor and trying to determine how they influence each other you leave all that for the neural network to figure out. It works - and it works well - if you keep within the boundaries for which it was made.

Could it be used for wargaming ? Well, let's say you limit it to the strategic level AI and a single problem, say Barbarossa on the corps/ army level. The developer then has to seed the network with lots and lots of "player does this, AI has to react thus" data. You immediately spot the problem with this : it's very hard to quantify the data. How would you define or determine that the Axis player has moved the weight of his armour south ? Influence mapping works in this case but then you've already done half the work outside of neural networks anyway. What should the AI do at this point ? Sheesh, move some units to the Don river line of course. You don't need a neural network to tell you that, in fact you (the developer) told the network that a zillion times.

Ok, say the developer has gone the extra mile with neural networks and allows the seeding to continue by the actions of the player - i.e. the "learning" ability of the neural network. So, ok, the neural network learns that every time the gamer puts Von Manstein on the Bucharest hex the gamer goes for the Southern option. Well, even if we skip over the problem that the network needs to recognize this as a factor (remember : only limited factors yield meaningful answers) then a
"defend against the Southern option" message may well be send to the AI. This works until the player recognizes this "programmed" response and adapts his tactics or (mis)uses it as a ruse or feint. In the end you end up with an AI that is not much better than what you could program yourself without this additional neural network stuff so you don't actually see it used a lot - if at all.

Greetz,

Eddy Sterckx

[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.)

Perhaps the most true of these three statements is #2 -- which is that neural networks need a lot of data to be trained. That is true if you don't provide scoring heuristics to help train the network. So, for example, I can have the network generate random input and score the output myself -- instead of relying on external data. But, that is often a difficult task and can lead to the network 'memorizing' the problem -- which is actually the biggest challenge with neural networks.

ORIGINAL: sterckxe
If you want to use neural networks for more general purpose problems or don't have accurate historical input AND output data or have a zillion factors to take into account for a single answer : forget it

Practically, that is absolutely true. However, I think that tends to be true of all AI in general. Deep Blue is one heck of a great chess player, but have it play Tin Soldiers, and it would get beat by any novice player.


At the end of the day, I think that there are a number of different approaches that you can take from an AI perspective to solve a problem. Some are well suited, some are not. Neural networks, as you say, have gotten a lot of buzz because they seem to work similarly to the human nervous system. But, we are mere cavemen in the exploration of the brain -- and trying to make a simple simulation of something so complex that you don't even understand is going to have major limitations.

(Just as an aside, we used the AI from Tin Soldiers Julius Caesar and used it to solve a business problem. It's for a company that schedules large oil tanker ships. We used the exact same algorithm and scheduled their fleet. Human time to create a new schedule was roughly 30-60 minutes. Our approach took about 2.5 seconds. And, using historical schedules, we demonstrated that it was always more profitable and on-time with its schedules than the human -- and it could be re-run constantly to attempt 'what-if' scenario planning.)

Deride

[Jeffrey H.]

ORIGINAL: Deride

... Another example I know is in the creation of random numbers. 'Truly' random numbers will still have streaks in them -- e.g., 1, 0, 1, ,1 ,1 ,1, 1, 1, 0, 1 ,1, 0. That string of 1's is entirely random, but players will see this and call the AI 'garbage.'
...

Did you ever try resetting the seed, to a sys clock based value every time the rand# routine is called ?

ORIGINAL: Deride

...(Just as an aside, we used the AI from Tin Soldiers Julius Caesar and used it to solve a business problem. It's for a company that schedules large oil tanker ships. We used the exact same algorithm and scheduled their fleet. Human time to create a new schedule was roughly 30-60 minutes. Our approach took about 2.5 seconds. And, using historical schedules, we demonstrated that it was always more profitable and on-time with its schedules than the human -- and it could be re-run constantly to attempt 'what-if' scenario planning.)...

Wow ! impressive. [&o] I'm currently working on a project that I hope will involve some AI, that way I can learn it for future work when I retire ! But it will be very simplified and not directly portable...or at least I htink it will....

Jeff

[Deride]

ORIGINAL: Jeffrey H.

ORIGINAL: Deride

... Another example I know is in the creation of random numbers. 'Truly' random numbers will still have streaks in them -- e.g., 1, 0, 1, ,1 ,1 ,1, 1, 1, 0, 1 ,1, 0. That string of 1's is entirely random, but players will see this and call the AI 'garbage.'
...

Did you ever try resetting the seed, to a sys clock based value every time the rand# routine is called ?

Jeff,

Actually, I wasn't talking about computer generated random numbers at all! I was just trying to point out that if I print out a list of random numbers (even pseudo-random which is the best computers can do), you will find streaks of numbers -- which humans, by their very nature, question.

I was trying to bring up the fact that artificial intelligence -- even if it isn't artificial -- is usually questioned when it behaves certain ways, even if a person were to do the same thing! Image playing a PBEM game against an AI opponent. Image how you respond to their moves calling it 'dumb' or 'smart.' Then, if I reveal, in fact, it had been a human the whole time -- your reaction would be entirely different.

(Certainly, we aren't there yet in terms of AI being indistinguishable from humans, but, at a certain point, I have no doubt that for games, the AI will be comparable -- and eventually better.)

Deride

[tomcat666]

To Deride

you mentioned that the AI in Tin Soldiers was very good, and utilized new techniques to make it unpredictable and challenging. so I look up reviews of the game on the Internet to get an unbiased assessment.

While there were a number saying it was capable there was this following review stating that the AI was predictable.

the review



"

Tin Soldiers: Julius Caesar
Computer Gaming World, September, 2005 by Di Luo

For a time, the Roman legions were the unquestioned masters of the Mediterranean world. In this latest addition to the Tin Soldiers series, you lead these storied legions against both barbarians and fellow Romans in the many battles of Julius Caesar’s career.
Most Popular Articles
in Arts

Tin Soldiers: Julius Caesar, like its predecessor Tin Soldiers: Alexander the Great, models itself after traditional tabletop miniatures war gaming. Its most unique feature is the simultaneous turn, in which both sides plot out their moves at the same time before executing them. Each turn is divided into three phases: main, reaction, and reserve. Orders are given to all units in the main phase, and depending on the result of those actions, you may get a chance to revise some of those orders later. This "we go" method adds unpredictability and suspense to an otherwise simple system. Even the best-laid plans will probably not survive contact with the enemy, especially if you’re playing against another human. And like Alexander, this new Tin Soldiers looks great, with units colorfully depicted as lovingly hand-painted miniatures. After decades of hexes and chits, it’s amazing how much life a little color and flair can bring to the staid war-game genre.

JC makes several improvements over Alexander. There are more unit types, including legions, which are more durable than traditional formations. The campaign is also less linear, allowing you to fight battles in different orders and giving you more control over the troops brought into each battle. Most importantly, the variety of battles has increased, including both defending and assaulting in a siege.


Unfortunately, JC kept many of Alexander’s faults. You still can’t play as Caesar’s enemies, except in multiplayer. THE A.I. REMAINS PRETTY PREDICTABLE, so most of the single-player challenge is due to the A.I.’s larger armies. Worst of all, there’s no PBEM. With battles lasting for hours, multiplayer games are just too difficult to coordinate and play.

Verdict

Good system and competent game crying out for PBEM."

So I ask again are there AI capability issues, that on some systems the AI operates well and on others it does not?

I w

[Zap]

Where are you Ravinhood? this discussion is just up your alley.

[ravinhood]

ORIGINAL: Deride

I had responded to the other AI-based thread that you posted as well. But, I think your question isn't really symetrical. Using the most 'modern' AI techniques is certainly not always going to produce the best AI.

For example, the standard pathfinding algorithm (A*) will often result in the optimum path to get from point A to point B, but players often will disagree with the path. Another example I know is in the creation of random numbers. 'Truly' random numbers will still have streaks in them -- e.g., 1, 0, 1, ,1 ,1 ,1, 1, 1, 0, 1 ,1, 0. That string of 1's is entirely random, but players will see this and call the AI 'garbage.'

I've worked with a number of AI algorithms and approaches (my Master's Degree was in the use of recurrent neural networks to solve human language problems.) In the end, the 'best AI', has to achieve goals that are beyond simply being 'smart'. They have to:

1) Meet the expectations of the user (for example, AI's that are too smart are often acused of cheating)
2) Adapt to the playstyle of many users (to be both a challenge and make the game possible to beat)
3) Be just difficult enough to put up a challenge -- but not hard enough to be frustrating (this is a *game* afterall)
4) Work on a wide group of machines within an acceptable performence watermark (which is HARD to accomplish with such variety of memory, processing power, etc.)

Trust me, I can always build an AI that can beat you at just about any game (Of course, my algorithm must be allowed to run as long as it wants whenver making any decision.) But, balancing the above 4 is what makes a truly great AI vs. a hard-to-beat AI.

Deride

Then why not add that feature as an optional highest difficulty because I for one don't think that you or anyone else can make an AI that can beat me without cheating (other than Big Blue). I start out all my games on the hardest difficulty and find the majority I can beat the first game at that setting. There are a "few" a very small few that I have to step down and play from a lower level to get better aquainted with how the game mechanics work against the ai. Once that is figured out I move back up to the hardest mode and stomp on the ai. So, I challenge you to make a game with this socalled "I can make an AI to beat you". Just make it optional, nobody has to play it if it frustrates them.

On the other hand I see a few that have hijacked this thread "Deride shame on you" getting completely away from "What wargame has the Best AI?"

Do we really want to hear all this mumbo jumbo of how AI's are built and neural networking and blah blah blah?

I'm more interested in seeing what people think about what IS the BEST AI RIGHT NOW, not about the history of the creation of AI's. lol

[tomcat666]

well ravinhoodd you must be the world champion of wargaming, because their are hundreds of others who find it very difficult to beat an AI opponent on the first sitting at the highest level.

But i do doubt the claim that this guy made about building an AI that can beat anybody. if so you should work for the US govt and built the terminator.

[tomcat666]

so what are you telling us ravinhood, only deep blue two can beat you in chess, I say bullshit.

I have a program called powerchess, that won the Harvard cup, beating the best players in the US, it beat 4 grandmasters and drew with two.

These guys were either us open champions or finalists.

I like to see you kick its ass.

what is your chess rating?

Sounds like a guy who is full of himself.

[ravinhood]

Ummm we are talking about computer wargame AI's, not about chess here. And I am talking about all the ones I have played in the past. Not a one is up to Big Blue's AI or the one you are talking about. So, when you him her can make an AI that can beat me in a wargame then we'll talk boy.

[jvgfanatic]

Oh wait, nevermind ^_^

[ravinhood]

Good apology accepted.

[jvgfanatic]

Well, you did mention Big Blue and thus you did bring up the idea of chess first (except for one earlier post that mentioned Chess earlier). If you want an apology I'm afraid you'll have to look inwardly for that (which should be no trouble given your proclivities). ^_^