Rerun of the Tromp-Taylor bet

[speedchase]

i think i can say for sure that for next 20 years...computer will not beat me.
computer go is improving rapidly but...it gets harder to improve one stone at this point.

In 20 years the top programs running on the best hardware will be able to beat anybody. And if the trend continues, that hardware will be Chinese.

Is this statement based on a specific article or something, or just your own observations? If its an article of some sort I'd like to read it if you could provide a link.

[Mike Novack]

It is hard to predict something like that. Remember that discovering the MCTS approach caused a jump in the playing strength of go playing programs compared to how the best regular AI programs were doing (progress with them had become rather slow).

We don't know but possibly the MCTS approach has a practical limit. Even great advances in crunch power might not help. You need to understand that if the leading cause of not selecting the strongest move comes from "sample size" (the actual probabilty vs the probability computed using the sample) this first improves rapidly as the sample size is really too small and then beyond a certain adequate sample size the improvement slows and slows as the sample size is increased.

A go game involves many moves so if we have a sampling size such that at the 95% confidence level about one in twenty moves won't have been the best choice. Even the 98% confidence level implies making a "mistake" or two in the course of the game. My point is that it can take a gigantic jump in sampling to get from 95% confidence to 98% confidence.

[yoyoma]

Only 1 hour left. Here are the current poll results, voting after the game starts is no fair!

Code: Select all

Who do you think will win?
Tromp 3-0 19%	[ 5 ]
Tromp 3-1  7%	[ 2 ]
Tromp 3-2 15%	[ 4 ]
Zen 3-0   15%	[ 4 ]
Zen 3-1   33%	[ 9 ]
Zen 3-2   11%	[ 3 ]

Zen leads the voting 16-11. I voted Tromp 3-2, but I still honestly have no clue, it could go 3-0 either way.

[Phelan]

It starts! And already very interesting at move 11.

[badukJr]

The way the game is going, Tromp should probably take stones so it doesn't get boring next games...

Kind of funny that he 'won' the bet playing against a laptop, then only a few months later it looks completely different.

[hyperpape]

Note that while one or two games in the previous challenge were on a laptop, the others were on EC2 instances that were about twice as fast.

[amnal]

And Tromp wins by resignation, congratulations to him.

[speedchase]

Tromp takes the first game by resignation. Chalk one up for Humanity ^^

[hyperpape]

If Tromp wins the whole thing, the next demand will be that he play in some tournaments so that the rest of us can assess his strength.

[Joaz Banbeck]

This is game 1. Sorry if the comments are a bit disjointed. I had to take out a lot to make it fit.

[Bill Spight]

Edit: Added four lost moves. Also deleted some comments to try to get the whole game to display.

[Joaz Banbeck]

That is better. Thanks. But B still has two moves in a row at 56 and 57. Anybody know where W played?

[yoyoma]

Computers are quickly getting strong at Go, but it looks like it is still very difficult to strip comments from SGF.

[daal]

Here's the game from gogameguru. It looks complete:

[speedchase]

does this work?


edit: ninjad