Game review (ELF)

[Gomoto]

ELF based self review

Steady win against a Fox 3D opponent. Never lost my fuseki lead in this one.
Some nice learning opportunities (aka mistakes by myself.)

[Gomoto]

What I learned from ELF this time:

Did not know this nice move at B6 yet.

[Gomoto]

Dont tenuki after kosumi:

(after low approach on 3-4)

[Bill Spight]

$$Bc Extension?
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . W . . |
$$ | . . . . . . . . . . . . . . . b . . . |
$$ | . . . . . . . . . . . . . . . a . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . O . . . . . , . . . . X , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------

Click Here To Show Diagram Code

[go]$$Bc Extension?
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . W . . |
$$ | . . . . . . . . . . . . . . . b . . . |
$$ | . . . . . . . . . . . . . . . a . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . O . . . . . , . . . . X , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Elf recommends the extension, . Old fashioned players often chose "a", and these days "b" is a popular local play. I am surprised they did not show up on Elf's radar. What does Elf say about them? Thanks.

[Uberdude]

I don't know about this position, but something I have noticed in my reviews with Elf and/or LZ is that they don't much like the "normal" 3 space extension after kosumi. The invasion point is an obvious weakness for later but I've generally made the 3-space as a "if pros do it so should I" kind of thing. The classic reason for 3 not 2-space is if kick and extend in corner then the extension is too close and black can pincer, the bots say this won't happen: if kick then generally tenuki and make another 2 space extension for a speedy development. There's a similar idea in approach 4-4, knight move, then 2 space extension: if kick then hane once and 2nd 2-space extension.

[Gomoto]

@Bill

a and b are not the best options according to ELF:

[Bill Spight]

Thanks. I'm surprised the differences are that big.

[dfan]

ELF in particular has very strong opinions.

It's a little bit of a quandary. The stronger a program is, the more seriously we should take its opinions (one would think). But the strongest possible program would just output 0% or 100% win probability for all moves, which would not be very useful when playing humans; playing a slightly non-optimal move in the opening would look just as bad as giving away a large group. Is there a sweet spot? Or is there a way to smooth the opinions of a super-strong engine so that they're more applicable to actual play? (Maybe if we can get engines to take winning margins into consideration, that would be sufficient. But so far that has been difficult.)

[Gomoto]

You can only decide move quality if you look into the continuations and variations.

And a move is only good if you can handle it.

(And than there it is also a question of style, often there more than one options. I for one dont like overplays and aggressive play. Personally I reach out to gain a little advantage with each of my moves and hope it adds up to a 2 point lead in the end . If I have the lead I try to simplify my game as much as possible.)

[Bill Spight]

You can only decide move quality if you look into the continuations and variations.

And a move is only good if you can handle it.

Pessimist!

[Bill Spight]

ELF in particular has very strong opinions.

It's a little bit of a quandary. The stronger a program is, the more seriously we should take its opinions (one would think). But the strongest possible program would just output 0% or 100% win probability for all moves, which would not be very useful when playing humans; playing a slightly non-optimal move in the opening would look just as bad as giving away a large group. Is there a sweet spot? Or is there a way to smooth the opinions of a super-strong engine so that they're more applicable to actual play? (Maybe if we can get engines to take winning margins into consideration, that would be sufficient. But so far that has been difficult.)

Well, as I have said before, today's top bots are not optimized for evaluation. Unless you can read the game out, which would give you 100% or 0% evaluations, since the margin of victory does not matter, you do not want to evaluate positions simply based upon how many points (on average) a player is ahead or behind. But you can add another parameter, the global temperature. You can estimate the final score by adding half the temperature to the current evaluation for the player with sente. Then you have an error term that is roughly proportional to the temperature, plus an error term that depends upon the strengths of the players.

For a bot optimized for play I doubt if estimating the margin of victory and the temperature would be more efficient than the MCTS approach, but I do not know if anybody has tried to do that. But I do think that the estimated ratio of the margin of victory and the temperature would make sense for humans. For instance, an estimated margin of victory of 3.5 pts. would mean one thing in the opening, with a temperature of 14, and another thing in the late endgame with a temperature of 2.

[mitsun]

A strong program should be able to estimate the winning advantage in terms of moves, which would be useful to humans. As a simple example, knowing my winning probability is 90% tells me little, but knowing I that I could pass and still win tells me a lot. (Of course this evaluation depends on current temperature being sufficiently low; even with a large lead, I cannot pass while playing out a ladder.) I think I heard that this is an experimental option for Leela, but I have not tried it out.