Slashdot Mirror
Chess: Man vs. Machine Debate Continues
Frederic Friedel sent in an interesting submission. It's an interview with the current world's chess champion, Vladimir Kramnik, in which they talk about the upcoming year in chess competitions, but also get into [Deep Blue] and where computer chess playing is versus several years ago, with a comparison between Deep Blue and Fritz. If you want more info, check out Chessbase for additional news.
It's been well known since, well, before I was born, that a computer could easily trounce a human in any game involving only tactics. For example, many fourth graders in this country have programmed a BASIC script to create a tic-tac-toe player that will never lose.
Therefore, it's not particularly novel that computers can beat people at tactical games. The only thing interesting that I see arising from these onging "human versus machine" chess matches is the proposition that strategy can be broken down into millions of tiny tactical evaluations.
This begs the question: is the strategy that a human chess player would use also based on these millions of tiny tactical evaluations, only so subtle that he's not aware they're going on in the vast electrochemistry of his brain? Or is strategy discernable from tactics in a human mind, but simply a subset thereof in a computer?
The sole interesting conclusion I draw is that if it can be proven that strategy is something different to man and machine, then a hybrid approach might allow us to solve problems in ways we've never dreamed of. Whether that hybrid approach would involve implanting computers in our minds, making computers that can function like minds, or simply working really well with computers, I leave to you.
"Beware he who would deny you access to information, for in his heart he deems himself your master."
Not necessarily. It is possible to write a program which knows only the rules of the game, and teaches itself how to play. This requires the programmer to be talented at writing machine learning code, but not necessarily talented at the game of chess.
An interesting page on the topic
This doesn't make much sense.
It's long been observed in AI circle that things that are seemingly difficult for human actually are quite easy for computer, and vice versa. E.g., it is relatively easy to write program to solve sophisticated equations, playing chess, etc, which usually are considered hard, and require long period of training for human to carry out adequately. Things that are easy for human, such as recognizing faces and doing common sense reasoning, are what present the most problem for AI researchers.
Turing test allows opportunity to test for the latter, greater challenge; your suggested test doesn't.
Certainly, given an infinite amount of processing power and storage space. But then, you can't just ring up Dell and say "Hi, I want an infitite amount of processors by next Tuesday"...And even if you could, it might not do you all that much good - You see, if you do a few calculations it turns out that there are more possible chess positions than there are atoms in the universe. Which might prove problematic when you are trying to store them.
Of course there are ways go about "solving" chess that don't require you to enumerate every possible board, but they are still way beyond the reach of classical computing, probably forever. Quantum computing might be a different story, but we'll just have to wait and see how that one pans out..
Say I construct a lookup table for every possible combination of moves. Then I eliminate every move which doesn't lead to my victory. I am left with a lookup table which contains the proper response to every move my opponent makes.
Horribly stupid idea. On average, there are 35 (?) legal moves from any position. Let's say we restrict a game to 100 moves, and we'll need a lookup table of 2.5*10^154 positions, which is obviously way more than than we could store if we used the entire universe for memory. Not to mention the time it would take to search all these positions.
Grandmasters can in fact tell whether their oponent is a computer, sometimes even after playing just a single game, and certainly by the end of a match. In fact, I believe Kasparov lost to Deep Blue precisely because he counted on the computeresque behavior of his opponent when designing his strategy. If you read the article, you will learn that Kramnik can tell computer programs apart by their style, and that he thinks Fritz is becoming more human-like in its behavior, from which I infer that he can still identify its style as computeresque on some level.
So, the test you propose has already been carried out, and the machines "failed". This may have more to do with the fact that the people who write chess playing programs are more concerned with the programs' ability to win than they are with the programs' ability to emulate the playing style of humans. If humans could calculate better [Note: "calculate" has a precise technical meaning in chess] or chess playing computer programs were slower and considerably more stateful, their respective styles might be much more similar and your test, therefore, be met.
My own belief is that the ability to play chess well, let alone the ability to play chess in the style of a particular grandmaster, is not an accurate or even adequate measure of intelligence, so I will not be particularly hurt when the day comes on which computers at last surpass our chess playing skills, just as they have surpassed our (numerical) computational skills.
Deep Blue II was composed of 2 frameloads of IBM SP/2 RS/6000 nodes interconnected by a proprietary crossbar switch. Each node had a specialized MCA-bus board which offloaded all automatable functions (move generation, position sorting...) freeing up the RISC processors to evaluate positions. The net result was that DBII could evaluate roughly 200 million positions per second. Deep Fritz 7.x on the other hand will run on an 8-processor Compaq Wintel machine and will be able to evaluate roughly 4 million positions per second.
The only wiggle room for making a reasonable comparison between these devices is provided by the assertion that the Fritz algorithms are so vastly superior to the Deep Blue II algorithms as to compensate for a difference of 2 orders of magnitude in computing power. This assertion is patently ridiculous.
Kasparov vs. Deep Blue II was a legitimate technological watershed. Kramnik vs. Fritz is a marketing effort by Chessbase GMbH. Period.
Even I can tell.
by the style of play, humans usually have clear strategies, computers dont, they usually just tactically try to beat you, using lots of tricks and traps, they dont have REAL plans so its easy to know its a computer if the computers every move is generic.
If you use Linux, please help development of Autopac
A clever argument, but completely wrong.
This argument would work only if you also knew every move the opponent was going to make. As long as the opponent is not completely predictable, this approach fails.
Let's get back to your decision tree: at any given moment, the trees branching off from each possible move probably contain *both* winning *and* losing outcomes. You can't "eliminate every move which doesn't lead to my victory", because there is no immediate next move which *always* leads to victory.
Ok, so you minimize the risk. Say, you count how many of the possible outcomes are losses and how many are wins for any particular move, and then you go with the move with the highest probability of winning.
Still doesn't work. A clever human may maneuver the game to one of the very rare losing outcomes for your hypothetical program.
That's part of the beauty of chess: there is nothing that one player can do to ensure victory. It all depends on the interaction between the two players, and that's what's been the hardest thing for computers / software / programmers to master.
Cheers
-b
Let me interpret for the original poster:
If you have a tree who's root is the current board position and the leaves are all "mate" or "draw" positions (where the mates give you the game) then it certainly WOULD be possible to force a draw or win for yourself in every game. Starting from some arbitrary board position, of course you couldn't, because there are then board positions which will not/can not happen. But FROM THE BEGINNING, a program could enumerate every possible move, and eliminate those which end in its loss.
Now, that said, it's not true that from a given board position, there is any move that will guarantee a loss. In some positions this is true, but not for any given position (take the first move of the game, for example. Presumably you can win no matter what first move you make.) The early game would likely need to be a series of moves to bring you to a board position in which the rest of the game is deterministic based on the tree. This is because eliminating all initial moves that can result in a loss will eliminate all initial moves.
Getting to that deterministic position is not guaranteed, either. If it was, the entire match would be deterministic based on the computer player's moves. So there would still be strategy involved. But the computer could still look ahead to prevent moves that result in mate (for example, the 4 move mate that's such a common ploy against new players). And in doing so, it could look for ways to get to a board position that's in its lookup table.