Depends on whether you want to win or lose.

Them's fightin' words!

_________________
Patience, grasshopper.

That may well be the explanation of why Tromp tried changing style. Presumably he knows about the behavior of this sort of algorithm.

a) Without anything being wrong with the algorithm the progamr can "make mistakes" (not play the best move). This is a matter of probability. Comparing the results of a large number of playouts following move A and move B, if A is the better move the percentage of victories will be higher than that of B. Usually, not always. Increasing the number of playouts will decreases the frequency of tis sort of error but that takes more time (or more crunch power for a given amount of time) and above a certain point the amount of improvement for each added playout becomes less and less (aka: diminishing returns).

b) The programs using a more straightforward AI might have weaknesses that could be learned. It appears that the programs using this sort of algortihm tend to play a "loose" style. Maybe that's a weakness compared to playing a territorially tight style but that's the sort of thing different pros would dispute.

You need to keep in mind that these programs do not always make the same move from the same position. Probability again. The closer the value of the second best move is to the best move the greater chance it would be selected.

And if one had little experience with humans who played this style preactice might be in order. But I wouldn't call that "learning a weakness".

Look at these games? If you didn't know which was the human player and which the computer would you have been able to tell? It might be interesting to conduct an experiment. We get presented with a number of game records and told that a third of them are between a human and a computer, a third beteen two humans, and a third between two programs and we see how successful we are at correctly identifying. Might start simpler with a number all known to be human vs computer but we aren't told whether the computer was playing white or black.

If you're thinking of people who have previously seen MCTS bots, the first three games would be pretty easy--they all have those characteristic floating center moves.

_________________
Occupy Babel!

The first three games were posted. Why not the fourth?

I don't know where they weren't posted, but all four were on the computer go mailing list and are now on the gogameguru site.

_________________
Occupy Babel!

i was curious how strong the program is and viewed few games.
even game i will crush him. 2 stone handy..i think i will win easily.
3 stone??? i am not sure but i think i will win.
i will play one game soon as i have some free time and let you know what happened.

_________________
"The more we think we know about
The greater the unknown"
Words by neil peart, music by geddy lee and alex lifeson

how do i play this Zen program...????
i need some instruction to play this Zen

_________________
"The more we think we know about
The greater the unknown"
Words by neil peart, music by geddy lee and alex lifeson

Please do

It would be very interesting to hear, and see, how you did against it.

@Magicwand: I haven't played zen for a few months now. Before what I did was to watch its current game and try to see when the opponent/zen might resign and then switch to the games list (before the game actually ends) and quickly try to accept the open game it comes up with next (it helps if you fan Zen and use your fan games list because you can save seconds from not having to find the newly-created game). Not sure about now, but a few months back the open game disappears sometimes, so you have to be prepared to immediately find the new open game if it does that.

Don't try to manipulate default handicap - you won't have time.

Look for Zen in the the computer go room on KGS

I do wonder if IBM could crack human Go superiority if Zen was ported to the Watson hardware.

For this challenge, Zen ran on 12 cores at 4.2 GHz. Watson has 720 cores at 3.5 GHz. Granted, the instruction sets differ between CPU architectures and there are going to be some challenges and decreasing efficiency at that networking scale, but if MCTS scales in a linear manner with computing power, that would make for a very interesting bout with some high Dan players.

It's definitely not linear. Improving the number of playouts by an order of magnitude seems to have a relatively small increase in strength for current MCTS bots.

_________________
Occupy Babel!

Monte Carlo engines have used supercomputers before, but not with any results that are much more impressive than this one. E.g in 2009, MoGo played against some professionals on the Huygens supercomputer, using 640 cores at 4.7 GHz. It managed to win one game out of four played on 7 stones against Zhou Junxun 9p.

Full list: http://www.computer-go.info/h-c/index.html

To give you an empirical sense of the non-linearity consider the results of what you can see right here at this site.

In the Markovich section we have some examples of fuego playing on a "standard machine" but given around 10 minutes per move rather than 10 seconds. On the same machine that is also a ratio of 60:1 and so comparable to what you suggest. In other words, that 60 fold increase in crunch power is worth a few stones at this point on the curve of fuego's playing strength. We shouldn't assume that another 60 fold increase would be worth a similar few stones. Likewise, as we decreased crunch power we would reach a place on the curve where the change of playing strength with crunch was very rapid (below a minimum amount, the algorithm would work very poorly if at all).

... Which makes complete sense when you think about human ranks as a log scale. As players, we are exponentially distributed in terms of skill, not linearly.

_________________
Tactics yes, Tact no...

The humans may have lost this battle, but I believe that on any hardware top professionals will kill the program in an even game. In 20 years, who knows who'll win. Humanity will eventually lose to the top programs.

_________________
My plan to become an SDK is here.

I sort of agree with the conclusion but not because of expected improvement in crunch power. Twenty years is a relatively long time in terms of the conceptual advances that have taken place in the programs. I don't expect that the MCTS programs will directly improve that much but we might have another breakthrough or a clever trick using MCTS to greater davantage.

Interesting replies.

Can somebody explain the jump in AI quality that's been achieved going from Deep Blue to the 2006 defeat of Kramnik against Fritz on what was quite modest hardware?

The crunch power won't hurt, though. It's going to have to be a combination of conceptual breakthroughs and computing power.



Deep Blue played really well in the match that it beat Garry Kasparov. Lest history evaluate that match incorrectly, Kasparov played much worse than usual, playing an anti-computer strategy, rather than playing normal and beating the machine.

Also, Kramnik's defeat in 2006 was kind of flukey. I say this because he blundered a mate in one in game 2, and then in game six he had to play a sharp tactical game, which played into Fritz's hands. Kramnik should have drawn that match.

In regards to the jump in AI quality, a number of techniques like LMR, have been invented, the evals are probably more accurate, and the overall speed of the engines have increased due to jumps in computing power. As of right now, chess engines are probably generally stronger than the best humans.

_________________
My plan to become an SDK is here.