does the AI "cheat" on the tech-trading in the end game?

[Ellestar]

Originally posted by duodecimal The point of this argument, if you can pay attention for a moment and concentrate, was a difference in opinion over what counted as "cheating". I was pointing out how scaling the intelligence of the AI could be used instead of giving the AI extra bonuses.

But you used a very bad example. Once again, GalCiv is exactly the same as Civ 4 - above average difficulty is achieved by giving AI extra bonuses.

[duodecimal]

True, no game (and no AI) can possibly compete with a good human strategist, so all games have to let the AI cheat with cheaper techs and improvements, and a head start at the beginning of the game. At least, for now.

There is no good example of a non-cheating AI, except in Chess. And all strategy games are much more complicated versions of Chess, so the problem is much worse. Chess is a good example of an AI that can not cheat: you can not grant extra moves or extra pieces, or allow more than one castleing, and if you want to upgrade a pawn, it has to get all the way to the other side. The AI and the Human in chess follow the exact same rules. This is not so in CIV, because the AI does get to cheat. Not as egregiously as before... but it ain't playing by the same rules and restrictions imposed on the Human.

I just wanted to be clear on what I consider 'cheating', and how I don't think it's accurate to say that the CIV AI (and the GalCiv AI) does not cheat at all.

[padillah]

Originally posted by duodecimal
True, no game (and no AI) can possibly compete with a good human strategist, so all games have to let the AI cheat with cheaper techs and improvements, and a head start at the beginning of the game. At least, for now.

I don't know that I agree with that. I like your comparison to chess and think it's perfect for this kind of thing.

For those that don't know, most chess games use a weighted decision tree of some sort. Given a certain board position or move series there are weights given to certain squares, captures, pieces and such. The "AI" searches through this tree to find the branch that leaves the biggest point gain.

The trick is the level of searching. If you search down a "sacrifice" line and see that by giving up your rook you could grab the others queen that should come up as a big point move. But if you stop at the sacrifice and don't search any farther you'll think it's a horrible move (cause all you see is loosing your rook) and it'll have 0 or maybe even negative points. (I've seen decision trees with point runs like +1, -2, -5, -14, -27, +60. It's a wierd thing to see).

You also make it so it doesn't search again for a number of turns. This keeps it from correcting itself once it realises it made the wrong choice and getting back on track.

This should be what the AI does in CIV. It has plans to build a big army or over develop culture, whatever the AI strat it has to recognize that moves are not singular events in time but are based on each other. So it can (and does) make "plans". Well, on Settler we could let the AI search the tree for maybe 16 levels every 8 turns; Nobel, maybe 30 levels every 4 turns; and Diety, maybe 60 levels every turn.

THAT'S how you cripple an AI.

It would also help give the different AI's personality traights that have more meaning. A "Cautious" civ might get +2-4 extra levels to think. An "Agressive" civ might take chances on negative, or "sacrifice", moves (lines that look negative but pay off in the end). An "Organised Civ might have a broader move base to search the tree... and so forth.
Tom P.

[tezster]

There's no way you can have anything resembling a chess search tree work on a game as complex as CIV. I'm sure you could apply a few broad principles, but as far as searching X amount of turns into the future... there are simply too many unkowns and random variables i.e. conditions change significantly from one turn to the next.

Search trees are best left to simpler games of complete information.

[Fried-Psitalon]

Sure, you could apply a chess-style search to CIV. If we were all running massive supercomputers and could actually spend three hours between each turn allowing the computer to compute that many cycles of possible computations. There are limits - and since people scream bloody murder when an AI's turn takes more than three breaths, there's an even bigger limitation.

CIV isn't chess. Not even close.

[duodecimal]

Unless you let the AI build decision trees during the player's turn. Set a few key player actions to cause the AI to recalculate its tree in between player turns, such as declarations of war, or changing strategic/health resource availabilities...

That ought to cut down on the amount of time spent between turns. The AI could be simulating its interactions while you're staring blankly at the screen, trying to figure out the best spot to put that settler.

I wonder how Gal Civ did that... the original one used multithreading on the OS/2 to let the AI plan its next move during the player's turn. I can't recall if the Windows version does the same (time between turns was very brief, so it might have been the case).

[DeepO]

Originally posted by jdchambe
Whether we call it cheating or not, the question remains the same:

Has any one else noticed that the AI's strategy for trading technologies in the end game changes relative to the human player only?

There are speific human-AI modifiers when it comes to techs, yes. This, to offset the ability of humans to actively create diplomatic alliances, while the AI merely follows the situation (which is partly random, partly created by the human)

However, this does not change in the end game. If a certain AI is unwilling to trade you techs in the early game, it will do the same in the end game. In the same situation. So, with the same tech-distribution (say, willing to give up a monopoly tech).

What you are seeing, is actually something a lot 'smarter': CIV AIs will consider whether a tech is directly leading to a vast advantage. It will not (easily) give away a tech with a crucial wonder it is building (and hence risks losing to you if it gives away the tech). This happens in the early game, but also in the end-game... the difference is that in the end game nearly all techs are critical as they will allow you to build a SS component, which you can use to win. It won't trade these component techs to other AIs either...

It destroys any sense that the perception numbers in the diplomacy screen mean anything.

Huh... the 'perception numbers' are a large part of the equation. They aren't everything when it comes to AI... it's a bit more complicated as that. Tokugawa is a typical example: he is an isolationist by nature, and whether he likes you are not, he won't ever give out techs that easily than say Ghandi or Mansa Musa. AIs are in a large part influenced by their personality, this just to avoid letting them look like idiotic numbers.

DeepO

[DeepO]

comparing chess to CIV is like comparing snooker with marbles.

Please remember that chess is the only example where a computer was able to beat a world champion. And it did this by sheer numbercrunching... the least 'smart' of the AI solutions. It was a show of hardware technology, which was just able to compute the outcome of so many more possible moves (23 in a row IIRC), so that human intelligence couldn't follow. Waiting for an AI to be able to do the same thing for CIV means waiting another 1000 years at least at the current improvement rate of hardware.

DeepO

[padillah]

Originally posted by tezster
There's no way you can have anything resembling a chess search tree work on a game as complex as CIV. I'm sure you could apply a few broad principles, but as far as searching X amount of turns into the future... there are simply too many unkowns and random variables i.e. conditions change significantly from one turn to the next.

Search trees are best left to simpler games of complete information.

I didn't mean for that to be a design document. I was just tossing that off the top of my head. I was outlining the way it was done in chess and positing a methodology for trying it in Civ. I never meant to compare the two games as if they were identical.

I don't even know how the AI works now, how could I possibly be expected to retro-fit an actual working emergent behaviour model? Does it think in terms of "what to do next turn"? Does it use a neural network to get better or does it have a set or rules to decide within? Does it use decision weighting at all or is there a predetermined set of steps for it to take?

I have no idea what is going on in there, why would I think I could fix it?

I just meant to propose a methodology, basically "something like that".

Tom P.

[jdchambe]

Originally posted by duodecimal
Fried answered yer question - he counts AI diplo bonuses as the kind of 'cheating' that's not in CIV. By his answer, the AIs treat each other no differently than they should treat you.

The AI is probably micromanaging its finances better than you are at this point, due to all the factors that are still being figured out.

Scan the logs to see how many Great Scientists they might have used to build Academies. That, combined with Free Religion civic, and other stuff can all add up, along with their traits that might be different from yours. Is this a single-game sample, or a trend you've noticed in several games? Are you playing at a higher difficulty level, where the AI gets to cheat a bit on research costs?

The AI will also refuse to trade techs that allow the building of Wonders that they may be constructing. Late game techs might be pretty heavily weighted with wonders and SS components.

fried did not specify whether they are treating me differently than each other in the end game when I read his response.

I understand that they don't want to trade resources that correspond to wonders that they are already working on. they always do that. i understand that this is why they are trading with me less in the end game also. its just that they seem to still be trading with each other to me.

as for the AI micromanaging better than I am: that is possible. I expect that this may be true for some of them. The reason I do not believe that is the only thing that is going on is because even the CIVs that have half the economic force that I do are having no problem rushing ahead of me on the research. I perfectly expect my nearest competitors to be able to micromanage the end game research better than I do. But what I'm having a hard time believing is that the smaller civilizations that had such a small ecomony that they were always multiple techs behind me leading up to the end game, are suddenly able to find some magic micro-managing formula that enables them to out produce me.

I'm fairly good at managing my economy to the point where my research is at 80%. Sometimes even 90% by the time the end game comes along afterall. It's easy to do if you can avoid war and keep your military relatively small. What I don't understand is how the small civilization right next to me that has twice the military infrastructure that I do and half the economy is able to out spend me on research, only in the end game.

[BgT]

Originally posted by DeepO
comparing chess to CIV is like comparing snooker with marbles.

Please remember that chess is the only example where a computer was able to beat a world champion. And it did this by sheer numbercrunching... the least 'smart' of the AI solutions. It was a show of hardware technology, which was just able to compute the outcome of so many more possible moves (23 in a row IIRC), so that human intelligence couldn't follow. Waiting for an AI to be able to do the same thing for CIV means waiting another 1000 years at least at the current improvement rate of hardware.

DeepO

I don't agree there. It is not numbercrunching. It had to RATE every possible move, or better said position on the board after x moves. These ratings are made with algorithms, which have to be good enough. That's why you have AI chess world championships - there are better and worse chess "programs". If we call numbercrunching combinatorics and the "other" style strategy, the AI must have good strategic algorithms to beat world champion. Combinatoric part was waaaay better years ago. Human player can rate two positions per second, AI can rate millions of them. However chess players still tend to choose openings that require lots of calculating because there is no way PC can do it all. For rating all positions after 5 moves it would take 40+ hours and for 10 moves 500+ years even for the fast chess computers. The 23 moves you mention is achieved with HEAVY selectioning best moves on every level. Done with strategic rating.

The numbers I wrote show, that Fried-Psitalon is unfortunatelly correct. There is no problem writing an AI that would always beat us in Civ - theoretically. But the time spent by AI would be forever.

[Apocalypse Cow]

Originally posted by duodecimal
There is no good example of a non-cheating AI, except in Chess. And all strategy games are much more complicated versions of Chess, so the problem is much worse. Chess is a good example of an AI that can not cheat: you can not grant extra moves or extra pieces, or allow more than one castleing, and if you want to upgrade a pawn, it has to get all the way to the other side.

Consider the AI to be an "average" Civ player who doesn't learn game to game, then translate the chess analogy. If someone only has average talent in chess, you can't expect them to beat a grand master, or even an expert. The only way you can get interesting games is to handicap the average player or the master, giving advantages or disadvantages in pieces. I've seen chess games where the master started out with 4 pawns, and 2 bishops, and he STILL won. From this light, the AI handicaps on higher levels make more sense.

It WOULD be interesting, however, for the AI to "take notes" on a user's playing style, tallying up over several games their tendencies. For example, there may be a file listing the proclivities of a certain profile, like "PlayerX has sneak attacked me 4 out of the last 5 times he's played, likes to beeline to judaism, and prefers military conquest to culture." That might make the AI more "human".

[Theseus]

Originally posted by Apocalypse Cow
It WOULD be interesting, however, for the AI to "take notes" on a user's playing style, tallying up over several games their tendencies. For example, there may be a file listing the proclivities of a certain profile, like "PlayerX has sneak attacked me 4 out of the last 5 times he's played, likes to beeline to judaism, and prefers military conquest to culture." That might make the AI more "human".

It's eventually going to be a lot more intense than that...

A friend of mine in IT security tells me that it is *already* generally assumed that the gaming industry will develop "shared learning AIs", whereby information about strategy, players, etc., will be dynamically shared and incorporated via an internet-based "AI backchannel".

From a security perspecitve, the obvious concern is that bad guys will somehow hijack same, and use it for nefarious purposes.

From a gaming perspective... WHOA!

[padillah]

You want to fry your brain? Take a look at some of these...

http://www.cs.stir.ac.uk/~lss/NNIntro/InvSlides.html

http://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html

http://www.statsoft.com/textbook/stneunet.html

... and finally here's one that sells neural networks as an aide to help your business grow.

http://www.nd.com/

I was researching them in order to build a "learning chess game". While I understand them, they are proving to be a bit better than I in implementation and it's not going the way I had planned.

A great game that does use ANN and AL is Creatures.

Tom P.

[DeepO]

Originally posted by BgT

I don't agree there. It is not numbercrunching. It had to RATE every possible move, or better said position on the board after x moves. These ratings are made with algorithms, which have to be good enough.

Deep Blue was vastly critised in the AI community (which I have been professionally part of for the past 8 years, BTW) because it did not use the best rating algorithms available, not even close. It was a very basic rating primarily based on how many pieces (weighted of course) were left after 23 turns. That's why I call it number crunching: instead of find good rating algorithms are good search strategies, it simply used the lowest possible, and calculated all combinations without any distinction. It didn't even feature A* search IIRC...

You can't use that technique in CIV, as your pc would probably be able to handle only 4 or 5 steps, and a typical game has a long term horizon of 100 steps. A good human will think 100 steps ahead, so you need some kind of AI that will try to do the same.

DeepO