lets review master

Uberdude · Post by **Uberdude** » Sun Dec 17, 2017 5:15 am

Bill Spight wrote: Well, I suppose that a 30% estimate corresponds roughly to being around 20-25 pts. behind (after ) by conventional human estimates. Out of curiosity, I evaluated that position with my simple, biased evaluation function, for an estimate of 26+ pts. for White. I usually figure the central bias at around 4 pts. at that stage of the game, which yields an estimate of around 22 pts. Surprise, surprise!

Why the difference? Short answer: Black is too concentrated and too low -- at least, by how my function figures things.

I wouldn't expect it to be anything as big as 20-25 points (presumably how much you expect white to win if two highly skilled players play out the game trying to maintain the point lead rather than reducing it for safety AG style -- wouldn't it be interesting if AG played as black against a white Ke Jie and we see if/how it can reverse the game!). Near the end of the game a 90% win can mean a almost certain half point win. In the WeiqiTV reviews we had comments around move 100-150 that AG now thought black was winning instead of even (I infer 55%-60%) and when pressed Fan Hui revealed that means just a half point win or so expected.

What happens if we calibrate your evaluation function (which I presume is some kind of influence/shadow map) by applying it to to the following position from AlphaGo self-play #3 (Redmond's review). I tried to find one with one player going more for influence, but I presume AG thinks the game is pretty even now (I doubt more than 5% departure from white's initial 53%).

Click Here To Show Diagram Code: [go]$$B $$ +---------------------------------------+ $$ | . . . O . . X . . . . . . . . . . . . | $$ | . . O X X X . X . . . . . O . . . . . | $$ | . O X X X O X . . . . . . X O O O . . | $$ | . . O O O O . . . , . . . . X X X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X . X . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . O . O . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . X . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . X O , . O . . . , . . . . . X X . . | $$ | . O . O . . . . . . . . X X X O O . . | $$ | . . . . . . . . . . . . . O O . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Bill Spight · Post by **Bill Spight** » Sun Dec 17, 2017 7:44 am

Uberdude wrote:
Bill Spight wrote: Well, I suppose that a 30% estimate corresponds roughly to being around 20-25 pts. behind (after ) by conventional human estimates. Out of curiosity, I evaluated that position with my simple, biased evaluation function, for an estimate of 26+ pts. for White. I usually figure the central bias at around 4 pts. at that stage of the game, which yields an estimate of around 22 pts. Surprise, surprise!

Why the difference? Short answer: Black is too concentrated and too low -- at least, by how my function figures things.
I wouldn't expect it to be anything as big as 20-25 points (presumably how much you expect white to win if two highly skilled players play out the game trying to maintain the point lead rather than reducing it for safety AG style -- wouldn't it be interesting if AG played as black against a white Ke Jie and we see if/how it can reverse the game!). Near the end of the game a 90% win can mean a almost certain half point win. In the WeiqiTV reviews we had comments around move 100-150 that AG now thought black was winning instead of even (I infer 55%-60%) and when pressed Fan Hui revealed that means just a half point win or so expected.

Well, yes, we don't really know what the 30% means. It is Master's estimate, and if Master played Master from that position my guess is that Black would win less than 20% of the time. We also have to remember that this early in the game the swings in point evaluation from move to move are quite large (as opposed to win rate estimates, which should not change much). Which again means that there is a good bit of uncertainty in the final result, even given optimal play. So 20-25 pts. was a guess, which I made before seeing what my influence function would say.

What happens if we calibrate your evaluation function (which I presume is some kind of influence/shadow map) by applying it to to the following position from AlphaGo self-play #3 (Redmond's review). I tried to find one with one player going more for influence, but I presume AG thinks the game is pretty even now (I doubt more than 5% departure from white's initial 53%).
$$B
$$ +---------------------------------------+
$$ | . . . O . . X . . . . . . . . . . . . |
$$ | . . O X X X . X . . . . . O . . . . . |
$$ | . O X X X O X . . . . . . X O O O . . |
$$ | . . O O O O . . . , . . . . X X X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . X . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . X . . . |
$$ | . . O . O . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . X O , . O . . . , . . . . . X X . . |
$$ | . O . O . . . . . . . . X X X O O . . |
$$ | . . . . . . . . . . . . . O O . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+
Click Here To Show Diagram Code
[go]$$B $$ +---------------------------------------+ $$ | . . . O . . X . . . . . . . . . . . . | $$ | . . O X X X . X . . . . . O . . . . . | $$ | . O X X X O X . . . . . . X O O O . . | $$ | . . O O O O . . . , . . . . X X X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X . X . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . O . O . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . X . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . X O , . O . . . , . . . . . X X . . | $$ | . O . O . . . . . . . . X X X O O . . | $$ | . . . . . . . . . . . . . O O . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Well, some 50 years after Zobrist came up with his influence function, nobody has found a hand-crafted evaluation function that produces strong play, and I don't expect to, either.

This is a position after

. My function estimates Black to be around 25 pts. ahead on the board. I think it is still appropriate to subtract 4 pts. for its bias, yielding a little more than 20 pts. But it is White's turn and White gets komi, so the estimate of the final score with optimal play should be in the neighborhood of 6 - 12 pts. for Black.

Uberdude · Post by **Uberdude** » Wed Dec 20, 2017 7:06 am

Next game for review is Master #36 vs Gu Zihao, the young Chinese pro who was 5p at the time but is now 9p having just won the Samsung cup. This had an interested trade following a wedge on the top side where Master says Gu made 2 big mistakes, and then a fairly but not super early 3-3 invasion which puzzled Redmond: why not normal approach, and AG says normal approach was a in fact a little better than 3-3, so it's not that normal approach was bad, but 3-3 was also possible and not bad either!. Full game is http://www.alphago-games.com/view/event ... 35/move/29, Redmond's review is https://www.youtube.com/watch?v=gWabLAkvpIQ.

Click Here To Show Diagram Code: [go]$$B AlphaGo Master (black) vs Gu Zihao 5p $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 3 . . . . . . 0 . . . . . . . . | $$ | . . . , . . . . . , . . . . . 1 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 5 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 7 . 9 . . . . . . . . . . . . . | $$ | . . 6 8 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 4 . . . . . , . . . . . 2 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

The game starts with Master making its characteristic big high shimari, as Redmond noted in his review (and me in my Master orthodox opening analysis) it prefers (by 1%) this when white's bottom left is a 4-4, and big low when it's a 3-4 (by 0.3%). Redmond thought this was because it wanted to approach the bottom left next, but in some example mainlines in the AG opening tool it actually 3-3 invades there. Gu then makes the solid knight's enclosure on the bottom left (best says AG) and Master shoulder hits. Gu pushes up (whereas in the same position Ke Jie crawled, -0.7%) and Master jumps, AG now says invading bottom right 3-3 is 2% better. Redmond did say it's a bit odd to spend a move on a centre move like this of uncertain value so early in the opening, and AG's 3-3 recommendation has clear value, but personally I think jump makes a more interesting game. Gu then splits the side, unsurprisingly AG thinks the approach is better (by 2.3%).

Click Here To Show Diagram Code: [go]$$Wm10 AG recommendation $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . . . 2 . 1 . b a . . | $$ | . . . , . . . . 3 , . . . . . X b . . | $$ | . . . . . . . . . . . . . . . d e . . | $$ | . . . X . . . . . . . . . . . f g . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . X . . . . . . . . . . . . . | $$ | . . O O . . . . . . . . . . . . 7 0 . | $$ | . . . . . . . . . . . . . . . 9 8 . . | $$ | . . . O . . . . . , . . . . . O 6 . . | $$ | . . . . . . . . . . . . . . . 5 4 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Interestingly it suggests a black one-space low pincer whereas if white approached for move 6 it suggests backing off with the knight's move as normal for AG: presumably as black has invested more in the upper left moyo with the shoulder hit and jump the pincer is relatively better (though we aren't told what AG thinks of knight move now, I suspect similar to the pincer). White's response to the pincer is odd: plopping a stone high on the side, I wonder if it views approach for pincer as a good exchange because you've created the option of an even more spacious 3-3 invasion at top right. AG doesn't show how white would deal with black continuing at the top side because it recommends black 3-3 invading at bottom right so loses sente and white does then 3-3 at top right (and just pushes with e-g instead of 2nd line hane connect, seeking sente). So what would white do if black didn't 3-3, here's a few of my ideas:

Click Here To Show Diagram Code: [go]$$Wm10 AG recommendation, my variation for 13 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . 5 . . . . . 2 . 1 . . . . . | $$ | . . . , . . . . 3 , . . . 4 . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . X . . . . . . . . . . . . . | $$ | . . O O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

If attach on top that weakens the top side stone, so extend on top to reinforce which also eyes thinness of the big high shimari? Feels weird though.

Click Here To Show Diagram Code: [go]$$Wm10 AG recommendation, my variation for 13 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . . . 2 . 1 . . . . . | $$ | . . . , . . . . 3 , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . . . . . . . . 4 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . X . . . . . . . . . . . . . | $$ | . . O O . . . . . . . . . . . . 5 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

If knight move to defend corner then make your own to stop the 3-3 invasion AG loves so much and keep territorial balance?

Click Here To Show Diagram Code: [go]$$Bm11 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . 1 . O . . 2 3 . . . . | $$ | . . . , . . . . . , . . . 4 . X . . . | $$ | . . . . . . . . . 6 . 7 . . . . . . . | $$ | . . . X . . . . . . . . . . . 5 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . a . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . X . . . . . . . . . . . . . | $$ | . . O O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Anyway, Master played checking extension from the left which AG says is 1.8% worse than 3-3 invading lower right (like a broken record) so it ends up about the same as if white had approached instead of split. AG really seems to think 3-3 invasions are first class opening moves in the corners that should come before moves on the sides. Approach, kick, extend all optimal says AG, Redmond noted that 15 could be low to be more territorial but Master was playing an attacking game and Gu added a move on the top with 16 which Master immediately attacked with the nice shape point of 17. Instead of this 16 (46.8% for black) AG recommends to immediately undercut the high move (45.1). AG's expected continuation below seems rather mild to me with the kosumi of 17, if Gu had known this is all AG wanted would he have more readily not defended his group? I suppose the undercutting move does make it harder to attack severely or make profit on the right.

Click Here To Show Diagram Code: [go]$$Wm16 AG's recommendation for 16 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . 4 . . . . | $$ | . . . X . . . . X . O . . O X . . . . | $$ | . . . , . . . . . 2 . . . O . X . . . | $$ | . . . . . . . . . . . 3 . . . . . . . | $$ | . . . X . . . . . . 6 . . . . X . . . | $$ | . . . . . . . . . . . . 7 . 5 . . . . | $$ | . . . . . . . . . . . . . . . . 1 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . X . . . . . . . . . . . . . | $$ | . . O O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

WIP

Life In 19x19

lets review master

Re: lets review master

Re: lets review master

Re: lets review master