pluggable AI

[shinygerbil]

An AI would theoretically never miss anything going on onscreen, no? That's a fairly large part of playing Defcon

[MegaDeathKill!]

But that's what is also known as bad AI! An AI that basically cheats is as bad as one that offers no challenge, you have to find a middle-ground.

[shinygerbil]

I was backing up KingAl's point that it is theoretically possible to design an AI that can beat a human..

[torig]

I was backing up KingAl's point that it is theoretically possible to design an AI that can beat a human..

A: Deep Blue.

A: Deeper Blue

A: Deepest Blue.

The question's already been given

(I haven't read it, but our mutual friend advised this: http://www.springerlink.com/content/u081887127044601/ "Deep Blue's contribution to AI")

[shinygerbil]

If/When we can make our own AI, mine's gonna be called "So Deep it's Practically Almost Black" Blue

[NeoThermic]

A: Deep Blue.

Not quite. The match was unfair, as Deep Blue had the opportunity to examine every game Garry Kasparov had played, where as Kasparov didn't have the opportunity to do the same. I also call on the fact that Garry asked for a re-match after he lost, and IBM declined. IF Deep Blue was as good as they said it is, then a re-match shouldn't be a problem, no?...

NeoThermic

[KingAl]

Mine, meanwhile, will be Deep Purple, and as such will Rock Hard™.

[torig]

A: Deep Blue.

Not quite. The match was unfair, as Deep Blue had the opportunity to examine every game Garry Kasparov had played, where as Kasparov didn't have the opportunity to do the same. I also call on the fact that Garry asked for a re-match after he lost, and IBM declined. IF Deep Blue was as good as they said it is, then a re-match shouldn't be a problem, no?...

NeoThermic

I thought the rematch occured, with the v2: Deeper Blue (instead of Deep Blue).
Also, why would it be unfair for the CPU to analyze all games played by Kasparov?

I thought -but I'm not an expert- there are only "so many" (workable) combinations in chess. Like "if my opponent starts off with a <insert name/move> it's best to counter with a <insert countermove>.
Kasparov knows this a/ theoretically b/ by vast experience, so what would be wrong with the computer being instructed on the same set of rules?
I know you're a (good) chess player, so do enlighten me

(No need for flaming ; it's hot in here as it is )

[NeoThermic]

I thought the rematch occured, with the v2: Deeper Blue (instead of Deep Blue).
Also, why would it be unfair for the CPU to analyze all games played by Kasparov?

No, it was unfair that IBM didn't allow Kasparov to analyze Deep Blue's games. What many people forget (or don't know, I'm not sure), is that there were two games. In 1996 Kasparov beat Deep Blue 4-2 in a decent setup. They upgraded Deep Blue (hence why people use the unofficial nickname of "Deeper Blue") for another match in 1997. It was this one that the computer won, 3.5 to 2.5; at this point Kasparov asked for one more match (best out of three, right? ), and was denied.

There's also a point in the second game of the 1997 match in which Garry played a nice computer trap. See, material gathering is a good objective of any computer playing chess, especially when the material can be obtained without loss of attacking pieces. Move 36b is the point in time at which the computer... did something unexpected. Here's the raw logs of the computer, calculating it's next move (in the above, the computer is white and next to move.)

Code: Select all

ash guess Pa6b5p,Guessing axb5
 8(4) #[Qb6](30)[Qb6](30) 30^ T=1
qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8b8 ra2a6 Qe8d8 pd5d6 Bc7b6
 8(6) #[Qb6](53)#################################### 53  T=4
qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8b8 ra2a6 Qe8d8 pd5d6 Bc7b6
 9(6) #[Qb6](53)#################################### 53  T=9
qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Rc8a8 kg1h2 Ra8b8 pg2g3 Qe8f8
10(6) #[Qb6](55)#################################### 55  T=33
qf2b6 Ra8a2r ra1a2R Bd6c7 qb6e6 Kg8h8 bc2e4 Qe8f8 kg1h1 Bc7d6 ra2a6 Rc8d8 ra6a7

The search depth is a bit strange, but the computer is calculating that the best move is the queen to b6. Typical computer. Move the attacking piece deep into the black side to take the pawn formation or other such ideas, it's a logical move, but in chess, sometimes logical moves are not the best. So, with logs like that, Qb6 is the next move...
white plays... Be4?!

The computer moved a move which it only looked at *once* in its logs:

Code: Select all

11(6)<ch> 'ab'
 #[Qb6](32)######## [Be4](37)############################ 37  T=182
bc2e4 Rc8b8 pg2g3 Qe8d8 ra2a6 Ra8a6r ra1a6R Bd6c7 ra6f6P
12(6)[TIMEOUT] 37  T=199
bc2e4 Rc8b8 pg2g3 Qe8d8 kg1g2 Ra8a2r ra1a2R Bd6c7 qf2a7 Bc7b6 qa7a6 Qd8d7

Why would the computer do that? It's an obscure and unexpected move, and everyone (including Garry himself) expected the computer to play Qb6. Hell, the commentators mention this themselves:

YASSER, that Qb6 looked like a better move.
Indeed we'll have to ask Deep Blue after the game why he didn't play this. Going back just a moment to really -- it would be a tragedy if for Deep Blue he doesn't win this game and misses Qb6. Did we find a defensive -- did he find a defensive resource -- if we just take the move Be4 back for just a moment and try to move Qb6.

MAURICE ASHLEY: Well, first of all we know that Deep Blue didn't "miss" Qb6. Because Deep Blue looks at every single possible move in the position and Qb6 obviously has concrete threats associated with it, so it didn't miss the move. It's possible that it just came up with a defense in its mind that was sufficient and we're going to have to search for it now to see what that might be. The other thing, too, though that was a good point made, is this idea that if a computer sees something it assumes that the human sees it.

YASSER SEIRAWAN: Indeed.

MAURICE ASHLEY: And although we might be terrified by Qb6, if it calculates ten moves from now and says, "Oh, no, you have this resource on move ten that equalizes the position," then it'll avoid that variation because it sees it and it assumes that you're intelligent enough to see it, too. And it will play some other move. In the meantime we might have resigned after Qb6, might have said oh, I'm dead, it's over, and it would have been better to play that move. So it's one of the weaknesses, one of the flaws of a computer that it assumes it's playing against somebody that's equally intelligent and therefore avoids these kinds of variations.

PATRICK WOLFF: Well, another important point I think is to understand how the computer thinks. The computer won't fail to see ways to win material within some short number of moves, but it will often misevaluate certain kinds of positions.

YASSER and I are very happy for white's position that this queen has managed to penetrate. Perhaps Deep Blue has a different evaluation. Maybe it's better, maybe it's worse, that this queen might be less-well-placed, so long as it doesn't actually find a way to win a pawn, it's going to be difficult for it sometimes to distinguish, as it's difficult for all of us to distinguish, what position we should really go for as opposed to another.

YASSER SEIRAWAN: Well, again, Patrick, I have a lot of problems with this, because the point is, if we just freeze for a moment this position in our mind, think about how nice white's queen is. It's penetrated black's camp. White has the a-file under control, there are immediate threats, and we compare to the game position that after Rxa2, you are suggesting -- and I agree with you -- that the countercapture Qxa2 -- you're going to get exactly the same position as Qb6, only the queen is already penetrated. The queen is landing on the a2 square. So it doesn't make any sense to me how Deep Blue has made the decision it's made, with Be4. After Bc7 then I just think -- no, I'll go for the endgame, Qe6+.

It's interesting to note at this point that no one expected the (Spanish inquisition) Be4 move at that point. At all...

NeoThermic

[KingAl]

I seem to recall a game Kasparov played against Deep Junior in which the latter stalled for half an hour...

As long as CPUs rely on conventional logic, they'll always have weaknesses - but I have no doubt that eventually they'll become insurmountable.

[torq]

I was backing up KingAl's point that it is theoretically possible to design an AI that can beat a human..

Even the present AI can beat some humans (or mouse clickers)

[shinygerbil]

I was backing up KingAl's point that it is theoretically possible to design an AI that can beat a human..

Even the present AI can beat some humans (or mouse clickers)

S...Wh...er...I er...I had no..I don't know what you mean! >_>

[NeoThermic]

I seem to recall a game Kasparov played against Deep Junior in which the latter stalled for half an hour...

As long as CPUs rely on conventional logic, they'll always have weaknesses - but I have no doubt that eventually they'll become insurmountable.

You can stall all you like, but in most professional matches the game is timed, and you automatically lose if you run out of said time.

NeoThermic

[KingAl]

Well, exactly - that's why I mentioned it: Kasparov completely flummoxed the thing. All such 'competitions' that we've recently seen are essentially biased in the computer's favour - where if anything the opposite ought to be true to ensure fairness.

[torq]

An AI would theoretically never miss anything going on onscreen, no? That's a fairly large part of playing Defcon

So? You can do the same. It's only a matter of attention. You wouldn't call this cheating, wouldn't you?

AI has the advantage of thinking fast and see everything at once but it is ruled by the strict logic and therefore his moves can be predicted and countered by a human. To the point: adding an open script based AI feature to Defcon won't ruin the game as many think. Writing a tough script is not an easy task even for a good programmer. Of course the best scripts would be distributed throughout the community but their drawbacks and weaknessess will soon be exposed and counter tactics will be developed. Besides, other scripts would appear. Imagine a contest when 6 different scripts play Defcon trying to outdo each other.