I thought I'd highlight a post by Hideki Kato on the computer go mailing list. It seems that some of the same ideas that have made Zen so strong actually make it less able to benefit from increased processing power. http://computer-go.org/pipermail/comput ... 04894.html

_________________
Occupy Babel!

While it's not directly stronger, it can play faster without drop in strength. Playing fast always put pressure to the opponent, even in normal or long time setting. Imagine Zen used less than a second to make a move. But in the end, what make Zen so strong is software which is still develop continuously. Hardware is really not the point here.

Also hideki is talking about massive parallel processing, which is different than raw processing power. A lot of computational problems become slow when divided between 100 s of cores. Single core flops may be different but he didn't say anything about it.

He speak specifically of parrallezing on cluster mode ie with adding computers. But the montecarlo algorithm scale well with the number of core in the same computer

badukJr: this is true, but I thought MCTS did pretty well in parallelization in the past. Also, aren't the biggest advances in computer speed coming from parallelization these days? (edit) Meaning that this would still slow down progress from more processing power.

Lex: what does he mean by "and the hardware" then?

_________________
Occupy Babel!

But that wouldn't be the case with this computatonal problem. If you turned a "monitor" on to see where the program was spending its time you'd see that would be in conducting the individual playouts. And each of these is a computation independent of any other playout and the tree grows from each node independent of its growth from other nodes.

This is a good sort of problem in the sense of the work being easily divided among parallel processers. Possibly even one "executive" managing the tree with a lot of "workers" doing the playouts it assigns them (that depends upon the ratio of the work involved in managing and then later evaluating the tree compared to the work of each playout). Once that limit were reached the problem remains easily subdivided with the "chief executive" assigning a portion of the tree (separate nodes close to the root) to different "assistant executives" each with their staff of "workers".

For many other sorts of computational problems not so nicely divided up.

He means additional computer