alphago zero

dhu163 · #1

It's a big day when we get 80 more self-play games from alphago. Although several of the games are alphago 30k level (600 move games!), it's still funny to see it grow up.

I've taken a look at two of the games so far, the Lee 001 only up to move 50 ish so far, not looking any further ahead yet.

We get to see alphago zero embarrass alphago master's moyo, and follow through with some pretty endgame moves and given up several points at the (filling dame at move 200 ish twice)

Several new ideas and tenukis, but we've gotten so used to it, it seems normal!

dhu163 · #2

My number one highlight is how alphago zero manages to live shockingly deep inside alphago master's moyos. What looks like committing suicide and throwing in more dead stones ends up leaving that tiny bit extra aji to live on the other side of the board!

Also I note the counter-example to the "capture 3 stones to make an eye" proverb. See the end of figure 5 game 11 to see that capturing 4 stones doesn't even make an eye.

Bill Spight · #3

dhu163 wrote:

Also I note the counter-example to the "capture 3 stones to make an eye" proverb. See the end of figure 5 game 11 to see that capturing 4 stones doesn't even make an eye.

I have a little trouble regarding that as a counter-example, since the marked area below is not an eye (assuming per convention that the Black stones are alive).

Click Here To Show Diagram Code: [go]$$ No eye $$ ----------------- $$ . . O C C C O . . $$ . X X O O O X X . $$ . . . . . . . . . $$ . . . . O . . . . $$ . . . . . . . . .[/go]

Yes, the proverb is imprecisely stated. So what else is new?

fycj · #4

Thanks

Uberdude · #5

I've been replaying the AG Lee vs Zero20 recently as I got rather bored of interminable 4-4 and 3-3 invasions. Something I noticed is AG Zero normally ends up with more corners than AG Lee. This reminded me of some analysis I did of my OGS games some years ago, in which I found, on average, I ended up with substantially more than half the corners, but less so in the games I won: I had something like a 90% win rate, and an average of >3 corners in games I won but only 2.5 in games I lost. So although me having >2 corners in all my games could just be down to a territorial style, the fact I had less corners in games I lost tended to suggest corners were valuable and a good way to beat me was not let me get them.

So here are the stats for the 20 AG Zero (20 blocks) vs AG Lee games. By having a corner I mean controlling the 2-2 and 3-3 points and having some territory there, so for edge case example top right in game 7 I said white has half the corner and black none as white has 3-3 and some territory but black has 2-2 and no territory. I also recorded if anyone got a big centre (which I classify as including or near tengen and > 25 points with some fudging).

Code:

#  AG Zero Colour   AG Zero corners    AG Lee corners    Big centre?
1  White            3                  1                 -
2  White            3                  1                 Lee
3  White            4                  0                 -
4  White            3                  1                 -
5  White            3                  1                 - 
6  White            4                  0                 -
7  White            3.5                0                 -   
8  White            3.5                0.5               -
9  White            3                  1                 -
10 White            3                  1                 -
11 Black            2                  2                 ~Lee
12 Black            3                  1                 -
13 Black            3                  1                 ~Zero
14 Black            4                  0                 -
15 Black            4                  0                 Lee
16 Black            2                  2                 -
17 Black            4                  0                 Lee
18 Black            3                  0                 -
19 Black            0                  2                 -
20 Black            3                  1                 -
======================================================================
Average (all)       3.05               0.78
Average (black)     2.8                0.65
Average (white)     3.3                0.9     

Also 5 of the 10 games with AG Zero as black start like this (a most common answer):

Click Here To Show Diagram Code: [go]$$B $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 4 . . . . . . . . . . . . 3 . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 6 . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 8 2 . . . . . , . . . 5 . , 1 . . | $$ | . . 7 9 0 . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

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

moha · #6

Uberdude wrote:

I've been replaying the AG Lee vs Zero20 recently as I got rather bored of interminable 4-4 and 3-3 invasions. Something I noticed is AG Zero normally ends up with more corners than AG Lee.

This would agree with the opinions that Zero plays more human-like than earlier versions. I take this as a sign of Zero's strength, where it recognises slightly more of the game (such as the value of corners). Earlier versions often played surprising moves, but I suspect at least some of them were actually not so good but a kind of a horizont effect (masked by it's still huge strength).

An important difference between Zero and predecessors is that Zero doesn't do MC rollouts. I think in practice this may be a step in the "less (evaluation) knowledge, more search" direction. The point is the tremendous speedup achieved this way (rollouts are very slow, even if AG used a third, reduced/faster network for policy there). If the evaluation quality doesn't drop too much when using the network only, the huge speedup can be directed to deeper searches (and the better the search, the less evaluation quality is necessary). Interestingly this was also tried in earlier versions but found not to give good results there - presumably the restructured and combined value network works better.

Anyway, this could be something that enables a more mature perception of the game. Seeing more Zero games would be nice.

alphago zero

Who is online