KataGo 7 komi self-play games

Interestingly, only 1 tie out of 8 games so far - and many more “dramatic” losses in terms of final score than most of my KataGo continuations of AlphaGo Teach positions.

4 Likes

Thanks! It’s cool to see how the alternatives play out.

2 Likes

Hmm, it’s kind of sad that the “improvement” is just avoiding a challenge rather than taking it on…

3 Likes

The release notes show, that they changed the selfplay-training as well to make the training more resilient to such blind spots.

Selfplay-training Changes

Added a new option to upweight training on positions with surprising evaluations, and a new option to help increase the proportion of “fairly” initialized games during selfplay.

6 Likes

Can you say messy and complicated? If you can, that means you’re equipped to describe game 9. :slight_smile: Half of the board turns into a giant semeai, and white wins by 76 points or so.

3 Likes

On Move 12, can we try C10 to overconcentrate Black?

2 Likes

Even though C10 is only KataGo’s 58th-favourite (!) move, it’s still only 6pp worse (45 % for white) than the game move (D6 in game 1 and 9, F3 in game 4, both 51 % for white). At this juncture of the game, there are still many possible moves. KataGo devoted 151 out of 911,313 playouts to C10. I’ve added the sequence. :slight_smile:

2 Likes

Game 10 was the first close game since the tie in game 3, white by 2 points it looks like. Peaceful game with huge territories on either side. Quite pleasing-looking to this 11k human :slight_smile:

3 Likes

Yeah, this game is easier to understand.

I’d be quite inclined to play #39 at R8, but I think that’s a style choice – running out through dame like Black does from then on is not really to my taste.

It’s also interesting that White plays the immediate connection on #28 instead of the “textbook” G17, because F13 has devalued the G16 cut.

1 Like

KataGo’s second choice. Black 50.5 % (-0.3 pp) at 50,370 playouts (out of 347,492).

Interestingly, by the time KataGo decides on 39. bQ6, the sequence it foresees is quite different from the “running through the dame” seen in the actual game.

Indeed, wG17 is only KataGo’s tenth-favourite move (-4 pp), but G18 as its second-favourite move seems to follow a similar idea (-0.3 pp at 139,695 playouts out of 274,714). Looks more complicated than J17 (white 48.5 %) to this 11k human.

Variations added :slight_smile:

2 Likes

Pleasant to know that R8 was barely considered inferior here. I play it a lot, since I generally consider it an integral part of the finished shape that one should aim for when attaching in the first place.

2 Likes

Game 11 (NZ rules, 7 komi) is back to messy and complicated. Very alien-looking to me. Black by 10 points or so.

2 Likes

Game 12 (NZ rules, 7 komi) is probably my favourite game in this thread so far. The fuseki looks like drunk humans playing, and after that it’s just one breathtaking magic trick after another - to end in only our second tie so far. Beautiful game.

3 Likes

At Move 18 it’d be nice to see what it thinks of the “normal responses” C6 and D6.

Then on Move 20, what it thinks of B4 (corner defence) and D6 (pivot towards the bottom).

I like that it does C8 rather than C9, seeing as Black has slide miai to the top and bottom.

On Move 36, what is the variation for Q6?

1 Like

D6 is gonna be game 13. :slight_smile: C6 is only its seventh choice (-0.5 pp).

B4 is its 13th choice (-3 pp). D6 is its 33rd choice (-10.5 pp).

C9 was its eighth choice (-1.5 pp).

Wasn’t sure if you meant 36. wQ6 or 37. bQ6, but KataGo likes both, so I’ve added both :slight_smile:

3 Likes

Split off game 13 (NZ rules, 7 komi) at move 19 rather than 18 as I said I would (added variations to make up for it). Huge territories on both sides, yet very close (looks like white by 2 points).

2 Likes

Game 14 (NZ rules, 7 komi) was a complicated game with small territories and a confusing endgame that ended up as our third tie.

Our current tally stands at 4 black wins, 3 ties, 7 white wins.

2 Likes

Game 15 (NZ rules, 7 komi) was another complicated game. A lot of small groups, including an early-game seki. The winrate stayed even throughout until black crashed at move 225.

2 Likes

I really need to make time to study these I feel. But I’m enjoying your summaries so much in the meantime! Thank you

1 Like

Game 16 (NZ rules, 7 komi) sees black going for max influence and building a giant center moyo. White grabs a huge bottom, sacrifices a group of 20 stones in exchange for the privilege to linearly reduce black’s moyo to nothing, and wins by 10 points or so.

1 Like