Union AI totally broken

[saintsup]

All I know (again, see my first post here - I was responding to a statement) is that Chess is much, much harder to play well than any computer game I've experienced. Magnitudes of order harder to play well.

IMHO, here is the heart of what is wrong or misleading in your opinion. Your statement above is literally true but ...

'Playing well' in chess now means mastering a game with 500 to 1000 years of existence, a potential player base of millions of people, playing during all their lifetime, from where you extract the best players and with some of the most clever minds of the world studying and writing books about the game.

OTOH playing AACW well now is being one of the best within the few hundreds of players who want to play it PBEM, with everyone with maximum 2 to 3 years of partial dedication to study the game and (almost) no literature.

It's obviously 'much harder to play well Chess' (let's say within the 10% best players of the world of chess) than to play well AACW(let's say within the 10% best players of the world of AACW). In fact with your standard of 'playing well' for Chess we all are playing VERY poorly at AACW.

You CANNOT deduce from this fact that Chess is more complex than AACW (in any meaning or definition of complex).

To try an other angle: Texas hold'em no limit nowadays is 'much harder to play well' than THNL 20 years ago. Why ? Because you have now millions of people playing it, including very clever/talented people and because you have tons of books that a lot of people are now reading/studying. But THNL nowadays is obviously not 'more complex' than THNL 20 years ago.

[MFogal]

Saintsup stated it better than I did: AACW is harder than chess, but there aren't hundreds of years and hundreds of books helping people get better. More to the point, it doesn't really matter which game is more difficult -- I enjoy playing both.

The reason it is interesting to look at chess is precisely because an awful lot of smart thinking has gone into programming computers to play chess. I'm not an expert programmer, but it seems to me that two things were done with chess:

1) Opening books: you can program the Ruy Lopez opening out to move 20+, meaning that the computer doesn't have to "think" in order to know what move to make. The equivalent in AACW might be to program the "Harper's Ferry gambit" where CSA takes Harpers Ferry with his first activated brigade, and then, if USA takes Manassas with more than 5000 troops. . . .

1a) Note: computers always beat chess players in "book" (unless the player is Kasparov, who wrote the book ...) That's because the computer can memorize moves easier than humans. Standard advice for a human playing a computer in chess: get out of "book" as soon as you can. Make a dumb move early, and you force the computer to think for itself.

1b) I'm not sure what the return to investment would be for AACW programmers to work seriously on "opening books."

2) Evaluation of a position: this is where chess computers have become much better in recent years. What makes a position good or bad? How can the computer improve its position? Is a sacrifice worth it (absent a clear, computable path to victory)?

2a) I think this is where the AI might be improved, and where the complaints about the AI seem to focus. For instance, about a month ago, Barksdale complained about an AI game where the computer sent a USA army at him without sufficient supply, which he was able to cut off and destroy. Questions: can the computer recognize that its troops are threatened? Can the computer calculate the moves Barksdale would need to take in order to cut off the AI army? Can the computer counteract the moves Barksdale would make to cut off the army? How does the computer place the reserve forces to make sure that the counter-countermoves are possible? etc.

2b) Instead of the brute force calculations described above, it's probably better to program some "heuristics" -- rules of thumb on how the AI will handle certain situations. I don't know if people agree with them, but the following are the types of rules-of-thumb I've got in mind.

Example1: every 10K troopers require either their own supply wagons, or are required to be within a set distance from a supply depot.

Example 2: the computer sets "strategic goals" every 3 months, and values those strategic goals more highly. (One trick I've used against the AI is to distract it from major cities by offering lesser cities. The computer "forgets" it was trying to take Richmond, and takes Charleston instead. This gives me time to rescue the capital.)

Example 3: mutual support is valuable, and the AI values a position more highly where MTSG is possible.

Example 4: unless there is an immediate goal, rest any body of troops (in a well-supplied region) that is down by 33% or more in cohesion. etc

I dont' want to argue with Granitestater about the relative merits of chess and AACW -- they're both good games. I would love to hear from more experienced AACW players about the rules of thumb that y'all use to evaluate positions. This might help improve future versions of the game. More importantly, I might improve my own play.

[Generalisimo]

I dont' want to argue with Granitestater about the relative merits of chess and AACW -- they're both good games. I would love to hear from more experienced AACW players about the rules of thumb that y'all use to evaluate positions. This might help improve future versions of the game. More importantly, I might improve my own play.

The general logic behind Athena (how we call the AI ) is common to ALL AGEOD games... because it is part of the AGE Engine.

Pocus tried to "insert" those "common behaviours" inside the AI... the problem is that there are so many variables at hand, that sometimes "good decisions" are overlapped by others "not so good", the AI makes a bet ... and voila, you have a problem.

That's why Pocus "liberated" some "AI guidelines" to the modding community so modders can help in the "polishing" of the AI behaviour, to help her (remember, it is her, Athena) do a better job.

This is not a "problem" in Chess, depending on how deep your decision tree is, the AI knows EXACTLY which are the posible outcomes of the player moves. So, that's why the AI can use the minimax rule, it just weights each decision and takes the "best" one from there.
BUT, only Deep Blue can process the entire decision tree... a PC can't do that, there is not enough computer power to do it.
So, each "difficulty" level that you select in a chess game is just the depth of the decision tree... usually, some years ago, it was like this (IIRC):
"easy" -> 2
"medium" -> 4
"hard" -> 8
On hard, the AI is being able to "see" what the player is going to do long before the player does the move. A normal player can HARDLY calculate 8 turns ahead like the AI can on expert... that's why the AI has the upper hand all the time on the game.

By the way, in the old days, you couldn't even run a decision tree with 3/4 depth... now you can (and I supose those figures have been update to new values with the Core2Duo processors and all that ).

[Mickey3D]

When you can model the "intelligence" needed to win a Chess game with a simple algorithm that uses Brute force and a simple decision rule like the Minimax, it clearly shows you that the complexity of the game is not that big.

Yes, chess needs a "crude" algorithm but because computers and humans are not "wired" the same way you are NOT modelling the human intelligence needed to win a chess game. Your are using the computer strength against a game that fit perfectly its capacity.

The problem is that a human brain just can't process all that info togheter and think in advance 20 turns, like a machine can while exploring all the posible solutions.
So, what does a player do to "win"?... think in advance X turns (2, 4, 10, etc depending on the ability of the player) and make the move knowing in advance the posible outcomes of that decision.

In AACW, that is just plain imposible. There is no way to explore posible solutions because the complexity of any type of solution to the game is just enormous when you look at the entire universe of posibilities that you have at hand.
So, what do you do?
Most of the time, you take decisions based on already proven success ... if you did X and win, you try to do X and win again...
That's a very simple strategy, but most of the people do it because there is no way to plan ahead the combinations of posible solutions to the "problem" you have in front of you.
So, because the problem is so big and complex, you just simplify it to take a decision.

Fully agree with you. But my point is that you can not judge the complexity of game from an human point of view based on the complexity required to solve it with a computer : you are comparing what is not comparable.


Perhaps this will help you understand what I mean (based on a story written long ago by a french author) :
The fox invites the stork to eat in a dish, the stork can't because of its long beak. Later the stork invites the fox to eat in a long vase : in turn the fox can't. What is more complex, to eat in a dish or in a vase ?

In case you are interested in the programmer point of view of this problem, a good and simple book that will give you an insight of how difficult it is

Working in the computer domain I fully understand the problem and agree with you that you'll need a more complex program in order to have the same level of expertise in AACW (i.e beating the masters) than in chess.

But, once again, my point is that you can't judge of the complexity of a game for a human player based on the complexity of the algorithm used to solve it.

even the problem to "decide which way to go", is "AI Techniques for Game Devlopment", by Mat Buckland.
It is not new, but it is one of the few that covers Genetic Algorithms and even Neural Networks.
There are many books around, but I found that this one is quite simple if you already have some background on the area.

I have a few ones on Genetic Algorithms, Neural Networks and fuzzy logic and I will add yours to my "to be read" list

+2

This should really be outside this subforum

+3

[Generalisimo]

Yes, chess needs a "crude" algorithm but because computers and humans are not "wired" the same way you are NOT modelling the human intelligence needed to win a chess game. Your are using the computer strength against a game that fit perfectly its capacity.

Fully agree with you. But my point is that you can not judge the complexity of game from an human point of view based on the complexity required to solve it with a computer : you are comparing what is not comparable.

Perhaps this will help you understand what I mean (based on a story written long ago by a french author) :
The fox invites the stork to eat in a dish, the stork can't because of its long beak. Later the stork invites the fox to eat in a long vase : in turn the fox can't. What is more complex, to eat in a dish or in a vase ?

Working in the computer domain I fully understand the problem and agree with you that you'll need a more complex program in order to have the same level of expertise in AACW (i.e beating the masters) than in chess.

But, once again, my point is that you can't judge of the complexity of a game for a human player based on the complexity of the algorithm used to solve it.

Well, I was not the one that started the comparison with Chess... but when you are trying to "emulate" the "human thinking" and you are trying to say which is simple/harder/complex/difficult, you need to use something to measure that statement.
Without any measure, there is no point in arguing... really.
I canl say that this forum is "pretty"... but what is pretty? what is the measure of pretty? is it pretty for all the people? is it pretty for you too? why not? it is pretty!!... and it is the neverending story...

I have a few ones on Genetic Algorithms, Neural Networks and fuzzy logic and I will add yours to my "to be read" list

+3

I too have many... but that book focus on the "games perspective" all the time... so it is really interesting... at least it was for me.

[GraniteStater]

Thank you, Mickey 3D; finally, someone who can grasp a distinction.

Generalisimo, maybe I wouldn't get so exasperated if I didn't read posts that almost seem to be a deliberate misconstruction of what is stated. I have gone out of my way to be crystal clear about, what should be to my mind, a fairly incontrovertible assertion, and I end up reading statements that lead me to question whether the author has really taken the time to understand what is written on the page.

Notice that I tried to be clear about my terms, as best as I could. I'm not a programmer and haven't done any serious mathematics in at least twenty years, but I do have a basic understanding of how these things work. Incidentally, I work in IT; I'm a technical writer and would like to think that I know how to express my ideas very well (I do - I've had many compliments from peers and editors and developers on the clarity of my work).

If anyone here thinks that AACW is harder than Chess - OMG, pass whatever you're smoking, 'cuz it's really good stuff. Are you freakin' kiddin' me? Don't tell me you actually have to think to play AACW. Are you serious? Yeah, there are some subtleties, but once you know what you're doing, c'mon it's an amusement, it's not a real challenge. Fun, but not an intellectual challenge.

Try to checkmate a lone King with King, Knight & Bishop against a random twelve year old who knows how to play Chess. Then come back and tell me how difficult AACW is.

"AACW is hard" - yeah, right. I haven't even played the CSA yet and within five games I'll be looking for ways to tweak it so I don't get bored.

[Mickey3D]

but when you are trying to "emulate" the "human thinking" and you are trying to say which is simple/harder/complex/difficult, you need to use something to measure that statement.
Without any measure, there is no point in arguing... really.
I canl say that this forum is "pretty"... but what is pretty? what is the measure of pretty? is it pretty for all the people? is it pretty for you too? why not? it is pretty!!... and it is the neverending story...

Once again, I do totally agree with you. But you are measuring the difficulty to emulate human thinking not the difficulty human perceives when playing the game.

Ok, I offer you to stop there or we could go like that indefinitely

[Generalisimo]

Once again, I do totally agree with you. But you are measuring the difficulty to emulate human thinking not the difficulty human perceives when playing the game.

Ok, I offer you to stop there or we could go like that indefinitely

hehe... sure...

Afterall, this is the AACW subforum, so we should focus on what we can do to improve the AACW AI behaviour...and in the end, Athena's behaviour.

[WhoCares]

When you can model the "intelligence" needed to win a Chess game with a simple algorithm that uses Brute force and a simple decision rule like the Minimax, it clearly shows you that the complexity of the game is not that big.

Yes, chess needs a "crude" algorithm but because computers and humans are not "wired" the same way you are NOT modelling the human intelligence needed to win a chess game. Your are using the computer strength against a game that fit perfectly its capacity.
...

Actually that's not totally true - as MFogal already said:

2) Evaluation of a position: this is where chess computers have become much better in recent years. What makes a position good or bad? How can the computer improve its position? Is a sacrifice worth it (absent a clear, computable path to victory)?

This is modelling the same that is done by human players. It is not like many people think that the best chess player is the one that can calculate furthest ahead. In most of the middle game it is more an evaluation of the current position and developing a plan where you want to get with your next moves. In that case it might be that he actually does not plan further than maybe two moves ahead, if at all...

This is also why computers can nowadays handle 'non-standard book' openings much better.

[Mickey3D]

Actually that's not totally true - as MFogal already

[...]

This is modelling the same that is done by human players. It is not like many people think that the best chess player is the one that can calculate furthest ahead. In most of the middle game it is more an evaluation of the current position and developing a plan where you want to get with your next moves. In that case it might be that he actually does not plan further than maybe two moves ahead, if at all...

This is also why computers can nowadays handle 'non-standard book' openings much better.

You are right, the heart of the game is the evaluation function. I was willingly simplifying my speach. But I believe the function that should be used by AACW to achieve the same degree of expertise a chess program is able today, would be more complex.

[kwhitehead]

I never said AACW was simple and avoided the term 'complex' as much as possible.

I did say that it's child's play next to Chess and will maintain that. I described one ruleset as being smaller than another's. Perhaps to some, appearances mean a lot, but don't be deceived by the elegant and spare and conclude that it's simplistic, or simple (not synonyms, BTW). AACW is not necessarily more complex than Chess just because it has more rules or more lines of code - I don't think you've been reading the posts or understanding them; I already covered that. I am distinguishing between games, as games, and the programming required for a particular program. I certainly hope that would be clear to the average reader.

Can't help it, got to jump in on this one.
AACW is more complex than Chess. Complexity isn't the same as difficulty to play or whether they make good games. Complexity is measured in the number of possibilities that must be considered in order to execute ones "move" and how many "moves" it requires to complete the game. Checkers is more complex than Tit-tat-toe because there are more possibilities for each move and more moves are required. Chess is far more complex than Checkers for the same reason. And, likewise AACW is more complex than Chess.

And think about something for a minute - you are asserting that a game (AACW - which could be replicated on your kitchen table with cardbourd counters) is soooo hard, that a human being for an opponent is going to give you a more challenging experience than the AI.

Whereas in Chess, the opposite is true.

Willya listen to yourself, for the love of Mike? Think.

I'd say that demonstrates that the game which is almost impossible to beat when played by a computer (i.e., shall we say, close to 'perfect', 'flawless' play) and is easier when played by mere humans, is more difficult to play well than the former case.

Chess is in no way simple but it is finite. That is why IBM and many others took it up as a challenge for testing the limits of a computer. It has a relatively small set of rules that are well defined and are executed on a very limited playing surface, 64x64. From the game point of view it is a near perfect game since both sides start equal with only the slight advantage to one side of going first which is easily compensated for by requiring a series of games played both as black and white.

It still took them 50 years to come up with a Master class program, a computer using 200 processors, and who knows how much money spent doing it.

The AI on games like AACW "suck" because no one is going to spend that kind of time developing them so a human opponent will always provide a better game experience than the AI. The AI is just convient for learning the game but is never a serious opponent unless to handicap yourself to give it a chance.

This doesn't mean that one couldn't develop an AI for AACW that would whip you ever time. The reason is AACW isn't a balanced gamed like Chess and it is so complex that you could develop a "killer" strategy for the AI that might win all or most games. This assumes that AACW has a solution. Most unbalanced games that try to simulate real world event do have solutions that always work.

That's it, I'm done. Feel free to assert there's little gnomes inside your computer, I don't care.

And everyone knows there are hampsters in there not gnomes.

[enf91]

This is what I think Granite is trying to say.
1. Chess is more complicated to play. This is probably the most-argued point of all, because AACW has so many more pieces and so many more ways to win and so many more regions. Chess, however, has EXACTLY one way to win, and a player has to figure out how to win IN THAT ONE WAY. Sure, the actual gameplay of AACW is more complicated because you might have a stack in California you forget about because it's so big. But is California really critical for victory? In fact, you could win just by enduring or by conquering/destroying your enemy. There are myriad ways to go about that, making that part complicated, yet at the same time simple. Choosing a specific path is not relevant (simple), even if actually executing it is. Chess has one way to win and, depending on how the rest of the game went on, might even be impossible to win. That is how he used "complex".

2. AACW is harder to program a computer for. Duh. It has to factor in so many different things that no computer on Earth, outside the human mind, could grasp them all. Plus, the game allows players to place new "pieces" and get more by event.

3. Chess is easier to play against a human, while AACW is easier to play against the AI. Chess is finite and small, so it can easily be managed by a computer. A human (generally) can't focus enough attention or memory on specific moves or contingencies, making him a weaker player. AACW, as the above point stated, is too complicated for a computer to play. As such, it is a weaker player.

I think that's everything. I'm not saying I agree or disagree, I'm just saying I see his point.

[Generalisimo]

...

I think we have already tried to explain in many ways that his assertion is wrong in many aspects... but right now, the discussion has reached a "point of no return", so at least for me, it is ended.
I really think we should let the discussion "Chess vs AACW" die and focus on more important things that will really help the AACW community... don't you think?

[enf91]

Agreed. I had a previous post saying "let's drop this", but no one read it. At least now I got a response.

[MFogal]

I agree that the chess vs AACW debate has reached a point of diminishing returns. More interesting: are there any "rules of thumb" that readers would suggest to either help Athena make decisions, or to help little ole me play the game better?

[Chaplain Lovejoy]

Thanks, GraniteStater, for giving me something (else) to think about. I'm a USCF lifer (rating: a lowly 1550) and an AACW fanatic. I wonder if the level of one's USCF rating correlates with how good an AACW player one would be. Hmmm.

[Gray_Lensman]

deleted

[Chaplain Lovejoy]

Would the AI get any better if there was an option available that allowed Athena to think for several hours before the player made another move?

[Gray_Lensman]

deleted

[Mickey3D]

As regards the AI and how it reacts (itemizes movement priorities), AACW performs a complex series of calculations at the end of each turn assigning values to locations based on a lot of things, (Pocus would have to actually list them all), but these calculations take into consideration things like objective status, whether the region is "owned" or not, etc. As a player, you have little if no real control over how these calculations are made other than to take control of defined objectives.

AI.ChgLocalInterest (Pocus quoted info):

<snip>

AI.SetAggro (Pocus quoted info):

<snip>

Thanks for this post that centralizes information related to AI tweaking commands.

Pocus has just made available a datadump file (recreated each turn) that shows the "AISelfVal/Local Interest" value from the preceding turn. (The LocalInterest value calculations are made at the end of each turn.)

How do you trigger the creation of this datadump file ?

hidden underneath is the necessity for who knows how many "conditional" statements to determine if it's appropriate at any point in time to maintain, change, or remove a "forced Local Interest" value of the specific regions as necessary. To my knowledge, this command is currently not being utilized

Could you give more information on this command (the way you did it for the AI commands) ?

[Gray_Lensman]

deleted