Life In 19x19

Schachus wrote:I doubt this has a more involved liberty count, because it needs to be done in preprocessing, before giving the position to the NN.

Yes, me too. My basic idea is simply that AlphaGo considers a basic liberty count at least on some level, whereas I often do not.

I played another game today.

I lost against KGS 1k, giving two stones.

Here's a review:

Highlights...

Position 1

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . X . . . . . . X O O . |
$$ | . . X , . . . . . , . . . X . X X O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . W . |
$$ | . . . X . . . . . . . . . . . . O X . |
$$ | . . . . . . . . . . . . . . . O . X . |
$$ | . . . X . . . . . O . . . . . , . . . |
$$ | . . . . . O . . . . . . . . O . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

The marked move above is probably too much. That's because, after his cut, etc., the cross-cut seems effective:

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . X . . . . . . X O O . |
$$ | . . X , . . . . . , . . . X . X X O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . 2 . |
$$ | . . . . . . . . . . . . . . . . 1 O . |
$$ | . . . X . . . . . . . . . . . . O X . |
$$ | . . . . . . . . . . . . . . . O . X . |
$$ | . . . X . . . . . O . . . . 5 , . . . |
$$ | . . . . . O . . . . . . . . O . X . . |
$$ | . . . . . . . . . . . . . 4 3 . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Instead, I should probably just pull back:

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . X . . . . . . X O O . |
$$ | . . X , . . . . . , . . . X . X X O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . W . . |
$$ | . . . X . . . . . . . . . . . . O X . |
$$ | . . . . . . . . . . . . . . . O . X . |
$$ | . . . X . . . . . O . . . . . , . . . |
$$ | . . . . . O . . . . . . . . O . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Might be OK.

Position 2

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . B . . . X . . . . . . X O O . |
$$ | . . X , O . . . . , . . . X . X X O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Not sure where to play after black's marked move. I wanted to try this:

Click Here To Show Diagram Code

[go]$$Wc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 2 X 1 . . X . . . . . . X O O . |
$$ | . . X , O . . . . , . . . X . X X O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 3 . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

But it just seems hard to make much of the group there.

Position 3

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X X X X . . X . . . . . . X O O . |
$$ | . . X O O O . . . , . . . X . X X O . |
$$ | . X X X O . . . . . . . . . . . . . . |
$$ | . O X O . . . . . . . . . . . . O . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X W . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Again, not sure about the marked move, above. I want to avoid getting heavy, like this:

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X X X X . . X . . . . . . X O O . |
$$ | . . X O O O . . . , . . . X . X X O . |
$$ | . X X X O . . . 5 . . . . . . . . . . |
$$ | . O X O 2 1 . . . . . . . . . . O . . |
$$ | . . O 4 W . . . . . . . . . . . . . . |
$$ | . . . 3 . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . . . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Position 4

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X X X X . . X . . . . . . X O O . |
$$ | . . X , . . X . . , . . . X . X X O . |
$$ | . X X X . X . . . . . . . . . . . . . |
$$ | . O X O X . . . . . . . . . . X O O . |
$$ | . . O O . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . X O . . . . . , . . . . W , . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . O . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

I think the marked move above is really greedy, and is the wrong idea.

Here, I think I should consolidate the bottom and right, and just fight with one group, like this:

Click Here To Show Diagram Code

[go]$$Wc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X X X X . . X . . . . . . X O O . |
$$ | . . X , . . X . . , . . . X . X X O . |
$$ | . X X X . X . . . . . . . . . 2 . . . |
$$ | . O X O X . . . . . . . . . 1 X O O . |
$$ | . . O O . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . X O . . . . . , . . . . . , . . . |
$$ | . . O . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . O . . . . . . . . . . . . . X O . |
$$ | . . . X . . . . . . . . . . . O O X . |
$$ | . . . . . . . . . . . . . . . O X X . |
$$ | . . . X . . . . . O . . . . O X O O . |
$$ | . . . . . O . . . . . . O . O X X X . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

What do you think?

Not studying go this week, but an idea came to me that I want to try, once I start up, again.

Comparing study to AlphaGo, I've come to the general idea that studying high level games (e.g. pro games) is a good way to improve one's policy network - when you see several examples of "good" play, it helps in identifying candidate moves during your own games.

A big question, then, is how I can more efficiently train my "value network" (assuming the same model for learning that AlphaGo uses - maybe humans don't optimally learn the same way as AlphaGo, but I'll just ignore that for now). So what does a good value network give you? The ability to accurately assess a position and identify if it's good for black or good for white. Based on past experience, I have a very biased value network - I have some amount of arrogance, and assume that, whatever the result, it's probably good for me. Okay, maybe not all of the time (e.g. when I die), but I generally have this type of optimism.

To train this, I'm thinking of taking notes while I play. Just a simple evaluation: black is ahead; or white is ahead; or it's about even. And do this at various stages of the game. I probably shouldn't use the chat window to take notes, but maybe I can do this on the side. Then after the game, I can compare my notes to some sort of review - perhaps with the help of analysis software that gives winning percentages, like CrazyStone (or Leela does this now, too, right?).

Maybe this can help me to overcome my bias. I'll try to remember this next time I play a game. Hopefully, I read this before posting another game in this thread, and I remember to try it out.

Note-to-self: Kirby, try this out before posting another game review here.

It seems that humans don't have a reliable equivalent of a "value network" - direct whole-board intuition about who is ahead. Instead, even the professionals have to count each time before reaching any conclusions.

Kirby wrote: Comparing study to AlphaGo, I've come to the general idea that studying high level games (e.g. pro games) is a good way to improve one's policy network - when you see several examples of "good" play, it helps in identifying candidate moves during your own games.

A big question, then, is how I can more efficiently train my "value network" (assuming the same model for learning that AlphaGo uses - maybe humans don't optimally learn the same way as AlphaGo, but I'll just ignore that for now). So what does a good value network give you? The ability to accurately assess a position and identify if it's good for black or good for white. Based on past experience, I have a very biased value network - I have some amount of arrogance, and assume that, whatever the result, it's probably good for me. Okay, maybe not all of the time (e.g. when I die), but I generally have this type of optimism.

To train this, I'm thinking of taking notes while I play. Just a simple evaluation: black is ahead; or white is ahead; or it's about even. And do this at various stages of the game. I probably shouldn't use the chat window to take notes, but maybe I can do this on the side. Then after the game, I can compare my notes to some sort of review - perhaps with the help of analysis software that gives winning percentages, like CrazyStone (or Leela does this now, too, right?).

Comparing notes is a good idea. However, you must remember that there is no such thing as the probability of winning the game, except 1 or 0, depending upon perfect play. All other probability estimates depend upon mistakes. So the question is, whose mistakes? My guess is that Monte Carlo estimates are no better than yours, or not much better, for players at your level. And they are probably worse than yours in the endgame. I do not know the basis of the probability estimates of AlphaGo's value network. Do they come from comparing AlphaGo vs. AlphaGo results in certain positions? Are they based upon AlphaGo's errors? In any event, it seems that some value in between the Monte Carlo estimate and the value network estimate gives the best results. Moi, I would want to know the estimates of the value network.

palapiku wrote:It seems that humans don't have a reliable equivalent of a "value network" - direct whole-board intuition about who is ahead. Instead, even the professionals have to count each time before reaching any conclusions.

How reliable? It certainly seems that AlphaGo's is better than that of current pros. But I expect that Go Seigen's was better than that of current pros, as well.

A secret that nobody seems to talk about is that the territory estimates that pros do do not produce an estimate of who is ahead. They have the practical value of giving an estimate of how much territory you have to make elsewhere in order to win. but that is not the same thing. Players like Takemiya, who rely upon influence, have to have a good sense of its value. As an influence player myself, counting territory early in the game would just make me depressed.

My sense is that pros sometimes make a quick assessment by counting bad moves rather than territory; e.g., "these three White stones look kind of inefficient in retrospect but everything else on the board looks pretty reasonable at a glance, so White must be behind." I'm not sure how relevant this technique is for weaker players, though; for one thing, we're even worse at evaluating moves than evaluating territory.

Another method I've seen is to evaluate the difference in points from a given exchange.

On an empty board, a joseki might be said to be equal for two players. So if one player gets countable profit, and the other gets influence, maybe you can say that the position is about equal. If one of the players make a questionable exchange, perhaps you can compare that to what would be more ideal and see the difference in points.

I agree that this method is somewhat mystical and perhaps prone to inaccuracies, but it's still interesting to me.

A few comments on your last game.



Main focus: Put Black to the test.

I have tried to take Alpha Go and apply it to my game, but it's really difficult. The value network idea is extremely difficult because there's so much to take into account. There's so many moves you don't end up seeing that could have been better. Glad to see you're trying this approach. I might try taking notes too. I play too much with feeling and I just play where I think is good, not necessarily because of some chain of logical reasoning, which I think is not good long term.

about alpha go style..

Lately, i decided to play what i think is a good move but was told that it bad by stronger players.
alpha go played it and professionals agreed but many are moves that was my first instinct.

as a result of my experiment i am losing more games than before... but sometimes those moves help me win.
i will continue experiment with this approach until my winning percentage increase...

i will let everyone know if i see improvement in near future.
it is my desprate approach to improve...since i am stuck at my level too long.

Coming back to go (knock on wood?).

I had a fun time at the go congress, and decided to join the AYD again in September. Hopefully, my family medical situation stays stable.

Anyway, I decided to play a game on my "Kirby" account on KGS, which is 1k.

Here's the game:



Compared to how I felt during the game, when I saw the final score, I was amazed it was so close - I felt like I was winning by more.

But going through the game, I see that I made a series of endgame mistakes - even ones that I can see.

Sadly, I'm not going to make all of the diagrams for this review. That's pretty time consuming compared to the benefit I seem to get for it.

I've been somewhat conscious of the benefit I get vs. effort I spend on stuff, lately.

Maybe it's part of getting older.

Both sides made endgame mistakes in the lower left corner. B199 feels good but actually helps W. The game would have been really close if B had turned the corner into seki.

Instead of playing a game today, I reviewed part of a professional game. I get a monthly Korean go magazine, so I just took a game from that. I looked through the game move by move, and sometimes looked at commentary.



The file above is only the partial game. It's getting late, so I couldn't finish going through it all. Next time I go over a pro game, I'll continue from that position.

Note: The game tree doesn't show the main line on the first branch. I moved through alternative variations on the main lines sometimes, so you have to dig around to find the real game. Look for the longest branch.