ROADRUNNER: World's Fastest Supercomputer

[Terry McCracken]

No question Kasparov could then and no doubt today if he put his mind to the task.

I'm certain Anand could as well. I just missed a win at G/5, playing the Black pieces over the net against Rybka 3 running on a Q6600 with 2gigs of hash the other day. I was very tired and mixed my move order which cost the game.

If I can get that beast into a lost position, at blitz with Black no less, then no question a top GM could as well and catch his limit!

What does G/5 stand for? Was it on Playchess?

Also computers can't play blindfold chess, only humans can.

Anyway any chess program is a great beauty regardless if it's stronger than the world champion or not and to write a strong chess program is a difficult task only relatively few people are capable of.

Funny, playchess is one of the few servers I haven't played on.

G/5=Game in Five Minutes.

[MattieShoes]

Engines distinguish between pawns quite well. Doubled pawns, backwards pawns, passed pawns, isolated pawns, blockaded pawns, weak pawns on open files, central pawns, flank pawns...

I think the problem here is that comps don't play like humans, so trying to use human thinking on their play can lead to wrong answers. Humans use something like a static evaluation of a position to decide who's winning and what course of action is best, then spend their time figuring out the implications. Comps rely mainly on an exhaustive search with a "good enough" static eval.

A strong human's static eval is much better than a computer's, but a computer's searching ability is much, much better than a human's. With a sufficiently deep search, computers figure out many things that aren't explicitly coded in their static eval. For instance, assume a comp that doesn't evaluate pawn structure at all, all pawns are worth a flat 100 points. If a passed pawn can promote within the horizon of the search, it doesn't need to recognize the worth of a passed pawn -- It'd happily sacrifice a 500 point rook to let the pawn promote to a 900 point queen, just as a human would.

So you can get any answer you want to. If you want to focus on evaluation of a position without searching, humans beat the pants off computers. If you want to focus on search, comps beat the pants off humans. If you look at results, it's still relatively close, but the best engines are obviously stronger than the best humans in most positions.

[towforce]

A strong human's static eval is much better than a computer's, but a computer's searching ability is much, much better than a human's. With a sufficiently deep search, computers figure out many things that aren't explicitly coded in their static eval.

Humans also do some looking-forward - the amount depending on how tactical the position is.

My rule of thumb is that excellent eval is approximately equivalent (overall - not in "particular" positions) to 5 ply of full-depth search (so a 7 ply search with poor eval would have approximately the same strength as a 2 ply search with excellent eval) - though I invite people with more experience in testing to put me right on what is just a guess.

[MattieShoes]

Hmm, that'd be interesting to test. Take crafty, remove eval and put in something simplistic like piece values and piece-square tables, and see how many ply deeper it has to search to achieve parity... hmm...

[Uri Blass]

No question Kasparov could then and no doubt today if he put his mind to the task.

I'm certain Anand could as well. I just missed a win at G/5, playing the Black pieces over the net against Rybka 3 running on a Q6600 with 2gigs of hash the other day. I was very tired and mixed my move order which cost the game.

If I can get that beast into a lost position, at blitz with Black no less, then no question a top GM could as well and catch his limit!

What does G/5 stand for? Was it on Playchess?

Also computers can't play blindfold chess, only humans can.

Anyway any chess program is a great beauty regardless if it's stronger than the world champion or not and to write a strong chess program is a difficult task only relatively few people are capable of.

This is not correct that computers cannot play blindfold chess.

computers always do it(they have no eyes to see the board and they only calculate)

Uri

[Dr.Wael Deeb]

No question Kasparov could then and no doubt today if he put his mind to the task.

I'm certain Anand could as well. I just missed a win at G/5, playing the Black pieces over the net against Rybka 3 running on a Q6600 with 2gigs of hash the other day. I was very tired and mixed my move order which cost the game.

If I can get that beast into a lost position, at blitz with Black no less, then no question a top GM could as well and catch his limit!

What does G/5 stand for? Was it on Playchess?

Also computers can't play blindfold chess, only humans can.

Anyway any chess program is a great beauty regardless if it's stronger than the world champion or not and to write a strong chess program is a difficult task only relatively few people are capable of.

This is not correct that computers cannot play blindfold chess.

computers always do it(they have no eyes to see the board and they only calculate)

Uri

I do agree here....this particular terminology can't be applied to the computer chess field....
Dr.D

[bob]

Hmm, that'd be interesting to test. Take crafty, remove eval and put in something simplistic like piece values and piece-square tables, and see how many ply deeper it has to search to achieve parity... hmm...

Time won't make up the gap, I can guarantee you that... However, it is an experiment I might try at some point in time, just to have a "number" for this question ("how many plies is the evaluation worth?")

[Uri Blass]

Hmm, that'd be interesting to test. Take crafty, remove eval and put in something simplistic like piece values and piece-square tables, and see how many ply deeper it has to search to achieve parity... hmm...

Time won't make up the gap, I can guarantee you that... However, it is an experiment I might try at some point in time, just to have a "number" for this question ("how many plies is the evaluation worth?")

There is no constant answer to it because with deeper search you may need more plies.

1 ply of normal evaluation may be even weaker than 2 plies of simple evaluation when 9 plies of normal evaluation may be stronger than 11 plies of simple evaluation.

Uri