Life In 19x19

Zen19D is 6d on KGS, and is doing decently at that level, winning 7 of 14 rated games so far. It's been less than a year since I made the "Computers reach 5d on KGS" thread, will the steady progress of computer Go slow down or stop anytime soon? Let's see your vote for when Computers reach 7d(!) on KGS.

Even Zen's weakest account on KGS, Zen19N, is 5d now. Zen19N plays on slower hardware and slower time settings than Zen19D, but it is still quite formidable!

Here is the last thread: http://www.lifein19x19.com/forum/viewto ... =18&t=3951

I don't think this is really "reaching 6d" after all, it is playing 9X15, which is the ideal timing for computer go. That being said, the further they go, the harder it will be for them to improve.

Maybe.

I think this all depends on the discovery of another "breakthrough" that might be able to address one or more of the outstanding weakenesses. It is entirely possible that the MCTS programs might be strengthened by some change to how the tree "grows".

Yeah, I think the most accurate measure is the performance of medium-slow playing bots, but fortunately, it seems that they move in line with each other, with a stone's difference or so in favor of the blitz bots.

And re: Mike, there does seem to be some possibility of a surprise upside, but that's not the only way that there can be advances. Right now the best bots have been steadily and (not-so) slowly improving through incremental changes. That could easily stop, but I think the best prediction is continued short-term improvement. Whether that can lead to sustained progress seems more dubious.

It's personally hard for me to wrap my head around a MCTS engine playing at strong professional level. Take Sakata's wrong-side peep as discussed in modern master games. If I recall correctly, it required a 21 move sequence to see the value of the peep. There are variants of Monte Carlo that are almost hybrids of alpha-beta (http://www.ru.is/faculty/yngvi/pdf/WinandsB10a.pdf), and I wonder if something more analogous to that is where we'll eventually have to go. The approach taken in that paper is obviously not applicable to Go, but perhaps the general idea of hybridization is applicable.

I'm pretty surprised there are so few people voting the longer periods, almost all the votes so far are for <1 or <2 years. Where did all the doubters go? I usually fall on the side expecting steady progress, but my own doubts are growing as we reach these higher and higher levels.

Zen19D has now won 62% of 58 rated games at the 6d level, quite a solid performance! It's offline now, probably getting equipment setup for the Takemiya vs Zen19 match. Thread about that match is here:
http://www.lifein19x19.com/forum/viewto ... =18&t=5572

Even taking 2 years would be a major slowdown in progress. I think a few years ago, a lot of us thought that programs would jump up into the dan ranks and then stall, but it just hasn't happened.

The time lag from a 5 dan fast bot 0+9x:15 to a 5 dan slow bot 20+5x:30 was 6 months. Zen19D reached 5 dan in June 2011. Zen19S reached 5 dan in January 2012. If the improvement stays steady then a slow bot 20+5x:30 will reach 7 dan in September 2012. I think the steady progress of 1 dan in 1 year is great. Even 1 dan in 2 years is good. In 2 years, a bot may be in the Top 100 KGS Players!

hyperpape wrote:Even taking 2 years would be a major slowdown in progress. I think a few years ago, a lot of us thought that programs would jump up into the dan ranks and then stall, but it just hasn't happened.

Ah the poll option intervals I made make it hard to know how if people think it will take slightly longer (e.g. 1 year, 2 months) or almost 2x longer (1 year, 11 monnths). I voted for <2 years, but that's only because I expect it to take slightly longer than 1 year.

Still I expected more people voting for the <3, <5, or >5 options.

Well it's been almost 4 years, and while Google's AlphaGo hasn't played publicly on KGS, I think it's safe to say that it's level is at least 7d KGS.

viewtopic.php?f=18&t=12644

5-0 against Fan Hui
494-1 against Zen/CrazyStone/Pachi in even games (I wonder which one beat it once)
99% against Pachi giving 4 stones
86% against Zen giving 4 stones
77% against CrazyStone giving 4 stones

Giving 4 stones to Zen and CrazyStone... The version of the software they have and the hardware they're using are probably KGS 5d... So you could even argue AlphaGo is KGS 9d...

But we really do need to put some kind of qualification on that in terms of "with how much computing power"

A program like Zen might be playing at 6d on kgs given a machine with about an order of magnitude more crunch power than the average home laptop. On our home machines it would play a few stones weaker*.

This latest report is of a program playing on a machine with an order of magnitude more than what Zen on KGS is using. While the MCTS algorithm scales poorly, not out of the question that it would play a stone or two better given an order of magnitude more crunch.


* The problem is the MCTS usually need modification to be optimized for the crunch power vs time control change. They typically have their pruning coded for particular values. So not really just a matter of changing the number of playouts used. Perhaps with the right changes to the pruning used, only a stone or two weaker. But without, there is danger of falling below the threshold where MCTS works well at all, and could be much worse.

Zen19X has just broken through to 7d KGS! If only Zen had achieved this 2 days earlier it would have been the biggest news in computer go in 4 years. Instead their accomplishment is somewhat overshadowed by Google's AlphaGo. Nevertheless they are the first program to do this publicly on KGS.

http://www.gokgs.com/gameArchives.jsp?u ... dAcounts=y

I update the senseis page on the history of KGS bot ratings (and also compressed it a bit since it's spanning almost 10 years now).
http://senseis.xmp.net/?KGSBotRatings

Also Zen19X is now the #45 player on the KGS top 100 list (Zen19 is #95, a few days ago it was the first bot to breakthrough to the top 100).
http://www.gokgs.com/top100.jsp

The vast majority of its games are moderate handicap against low dans, with only one even game against a 7d which it won (unsurprisingly by a big centre kill against a human who underestimated its centre killing ability). I think we need more even games to say it is 7d with confidence (problem is the scarcity of KGS 7ds, and if they are online probably the larger number of weaker players will snipe the game offers).

Zen19X just beat 2 6ds, 2 7ds, and 2 8ds (inc. cheater = Pavol Lisy Euro 7d/1p) so that 7d is looking more believable now. Its play is impressive. The game vs Pavol:

Uberdude wrote:Zen19X just beat 2 6ds, 2 7ds, and 2 8ds (inc. cheater = Pavol Lisy Euro 7d/1p) so that 7d is looking more believable now. Its play is impressive. The game vs Pavol:

http://files.gokgs.com/games/2016/1/31/cheater-Zen19X.sgf

There is no komi however so Zen had a handicap of some sort.

Winning against european pro with komi handicap is a great improvement however AlphaGo can do so much more

Props to the Zen folks though; it's running on *way* less hardware (2x12 core Xeon server).