LeelaZero's "under the stones" strategy

sorin · Post by **sorin** » Thu May 24, 2018 2:00 pm

This is a training game from LeelaZero from yesterday.
It lasted for 623 moves!
What I found even more interesting was LeelaZero's "under the stones" strategy with black - see move 365 for instance, where instead of capturing two stones with P9 (which is not enough to win for black), it fills in a point, lets white capture that huge group, then later starts playing in the empty area created by the capture in an attempt to narrow the gap

Black does that several times in fact, with multiple groups, ending up with all the stones dead.

Somebody should tell LeelaZero that this strategy doesn't work

Tryss · Post by **Tryss** » Thu May 24, 2018 2:07 pm

Pretty normal. I mean, black is not allowed to resign, and passing when you're behind is the worse possible move, so it makes sense black try stupid things.

sorin · Post by **sorin** » Thu May 24, 2018 3:08 pm

Tryss wrote:Pretty normal. I mean, black is not allowed to resign, and passing when you're behind is the worse possible move, so it makes sense black try stupid things.

I see - but why doesn't white pass earlier?

Bill Spight · Post by **Bill Spight** » Thu May 24, 2018 3:56 pm

sorin wrote:This is a training game from LeelaZero from yesterday.
It lasted for 623 moves!
What I found even more interesting was LeelaZero's "under the stones" strategy with black - see move 365 for instance, where instead of capturing two stones with P9 (which is not enough to win for black), it fills in a point, lets white capture that huge group, then later starts playing in the empty area created by the capture in an attempt to narrow the gap

Black does that several times in fact, with multiple groups, ending up with all the stones dead.

Somebody should tell LeelaZero that this strategy doesn't work

Given the playing level of LeelaZero, I find it hard to believe that this is a new phenomenon. It sounds like what Freud might call the return of the repressed.

It may be an example of the loss of learning during a time of drift, in which successive winners are not actually improvements. It is not very serious, as Black simply lost a lost game. And I expect that LeelaZero will soon learn that Black's strategy does not actually increase already remote chances of winning.

Tryss · Post by **Tryss** » Thu May 24, 2018 4:15 pm

And I expect that LeelaZero will soon learn that Black's strategy does not actually increase already remote chances of winning.

No chance.

I mean, if Black pass, its winrate is exactly 0%. If black has 0.000001% chance to win by playing, then it's logical to play. LZ has no reason to learn to pass like we humans do.

Here in this training game (less than 7 days ago) LZ only passed because it didn't have a legal move available

Bill Spight · Post by **Bill Spight** » Thu May 24, 2018 4:52 pm

Tryss wrote:
And I expect that LeelaZero will soon learn that Black's strategy does not actually increase already remote chances of winning.
No chance.

I mean, if Black pass, its winrate is exactly 0%. If black has 0.000001% chance to win by playing, then it's logical to play. LZ has no reason to learn to pass like we humans do.

You mean that as long as there is one win in its playouts, the player who is behind will not pass?

But the players can pass, so presumably there may be a playout that begins with a pass, yet the player still wins. So we cannot say that if Black passes its win rate is 0.

Here in this training game (less than 7 days ago) {sgf omitted} LZ only passed because it didn't have a legal move available

Well, in the game shown by sorin Black passed with legal moves available at move 421, without committing seppuku. As sorin says,

sarin wrote:but why doesn't white pass earlier?

I.e., at move 422.

Has LeelaZero always played like this, or is it just that sorin did not notice until now?

Bill Spight · Post by **Bill Spight** » Thu May 24, 2018 5:51 pm

Click Here To Show Diagram Code: [go]$$Wcm20 Not a mistake? $$ --------------------------------------- $$ | X . X O O O O O O O . O O . O O . O . | $$ | . X X X O . O O X X O O O O O O O . O | $$ | X . X O O O O O O X X O O O O O O O O | $$ | . X X O X O O O O X X O O O O O . O O | $$ | X X X O X X X X X . X X O O O O X . X | $$ | X X O X X X O O X X . O O O O X X . O | $$ | X O O O X O O O X X X X O X X O O X O | $$ | O 1 O a O O O O X X X X O O X O O O O | $$ | O O O O X O . O X X O O O O . . O . O | $$ | O O O X X X O O X O . O X . O X X O O | $$ | O X O X . X X X O O X O . O O O O . O | $$ | X X X X X O O O O O O O X X O O . O O | $$ | X X X X . X O O O O O . X O O . X O O | $$ | X X X . X X O O O O O . X O . O X O O | $$ | X X . X X O O O . O O O O . O O O . O | $$ | X X X X X O O O . , O . . O . O O O O | $$ | X X X X O O O O O O O O . O X O X X X | $$ | X X X X O X O . O . O O O O O O O X . | $$ | X X X X X X O O . O O O . O . O X X X | $$ ---------------------------------------[/go]

White 420 commits seppuku. Sure, White can still win. But can anybody claim with a straight face that filling an eye needed for life increases White's chances of winning in any meaningful way?

Then Black 421 passes.

Does that give her better chances than capturing at "a"?

Click Here To Show Diagram Code: [go]$$Bcm29 Not mistakes? Really? $$ --------------------------------------- $$ | X . X O O O O O O O 4 O O . O O . O . | $$ | . X X X O . O O X X O O O O O O O . O | $$ | X . X O O O O O O X X O O O O O O O O | $$ | . X X O X O O O O X X O O O O O . O O | $$ | X X X O X X X X X 1 X X O O O O X . X | $$ | X X O X X X O O X X . O O O O X X . O | $$ | X O O O X O O O X X X X O X . O O X O | $$ | O O O a O O O O X X X X O O . O O O O | $$ | O O O O X O . O X X O O O O O . O . O | $$ | O O O X X X O O X O . O X . O X X O O | $$ | O X O X . X X X O O X O O O O O O . O | $$ | X X X X X O O O O O O O X X O O . O O | $$ | X X X X . X O O O O O . X O O . X O O | $$ | X X X . X X O O O O O 3 X O . O X O O | $$ | X X . X X O O O . O O O O . O O O . O | $$ | X X X X X O O O X X O . . O . O O O O | $$ | X X X X O O O O O O O O . O X O . . . | $$ | X X X X O X O . O . O O O O O O O . O | $$ | X X X X X X O O . O O O . O . O . . . | $$ ---------------------------------------[/go]

= pass

Later, Black 429 puts her head on the chopping block instead of capturing at "a". White 430 does not deliver the coup de grace, but passes.

{no longer shocked}. Then both players continue to avoid capturing the opponent's large group while saving their own. White finally captured the Black group at move 450.

Does no one think that Leela Zero has things to learn here?

dfan · Post by **dfan** » Thu May 24, 2018 6:09 pm

Bill Spight wrote: Has LeelaZero always played like this, or is it just that sorin did not notice until now?

My impression is that Leela Zero was more reasonable about passing until it started learning from the games of ELF, who passes a lot less (my understanding is that the ELF code handled passing specially, while Leela Zero learned it naturally in its neural network).

Bill Spight · Post by **Bill Spight** » Thu May 24, 2018 6:17 pm

dfan wrote:
Bill Spight wrote: Has LeelaZero always played like this, or is it just that sorin did not notice until now?
My impression is that Leela Zero was more reasonable about passing until it started learning from the games of ELF, who passes a lot less (my understanding is that the ELF code handled passing specially, while Leela Zero learned it naturally in its neural network).

Thanks.

ez4u · Post by **ez4u** » Thu May 24, 2018 6:21 pm

Discussion here: https://github.com/gcp/leela-zero/issues/1446

Life In 19x19

LeelaZero's "under the stones" strategy

LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy

Re: LeelaZero's "under the stones" strategy