Kirby's Study Journal

Bill Spight · Post by **Bill Spight** » Fri Aug 24, 2018 1:22 am

Click Here To Show Diagram Code: [go]$$Bc Which is better? $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . b . . . . . | $$ | . . O O . O . . . . . 3 1 O . . . . . | $$ | . . . X . . . O . , . . 2 a . O . O . | $$ | . X X X . X . . . . . . . . . . . . . | $$ | . O O X O . . . . . . . . . X . X . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . X X . . | $$ | . O . . . . . . . . . . . . . X O . . | $$ | . . . X . . . . . , . . . . . , O . . | $$ | . X . . . X . . . . . . . O . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

attaches, to make sabaki. After

the question arises, which is better, "a" or "b"?

Click Here To Show Diagram Code: [go]$$Bc Maybe neither. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O O . O . . . . . . 1 O . . . . . | $$ | . . . X . . . O . , . . . 2 . O . O . | $$ | . X X X . X . . . . . . . . . . . . . | $$ | . O O X O . . . . . . . . . X . X . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . X X . . | $$ | . O . . . . . . . . . . . . . X O . . | $$ | . . . X . . . . . , . . . . . , O . . | $$ | . X . . . X . . . . . . . O . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

My feeling is that the hane strengthens Black unnecessarily. Instead, how about the stand? Now you don't face that problem.

Click Here To Show Diagram Code: [go]$$Bc One possibility. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . 6 . . . . 3 4 . . . . | $$ | . . O O . O . . . 5 . . 1 O . . . . . | $$ | . . . X . . . O . , . . . 2 . O . O . | $$ | . X X X . X . . . . . . . . . . . . . | $$ | . O O X O . . . . . . . . . X . X . . | $$ | . . X O . . . . . . . . . . . . . . . | $$ | . . X . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . X X . . | $$ | . O . . . . . . . . . . . . . X O . . | $$ | . . . X . . . . . , . . . . . , O . . | $$ | . X . . . X . . . . . . . O . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

In this variation,

is met with the footsweep.

Gomoto · Post by **Gomoto** » Fri Aug 24, 2018 1:29 am

KISS works quite well in the fuseki:

KISS option 1:
This is never a bad move in the fuseki

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

KISS option 2:
The shimari is bigger than a possible black double approach.

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

Both KISS options work quite well.

I prefer option 2 because black did not follow this KISS fuseki rule: An approach to 3-4 is bigger than an approach to 4-4

(I personally follow a even more strict KISS fuseki rule: Is the low approach to 3-4 available? Take it!!!)

explo · Post by **explo** » Fri Aug 24, 2018 2:06 am

What is KISS?

Uberdude · Post by **Uberdude** » Fri Aug 24, 2018 2:29 am

Here's what I'd play these days, having embraced the idea from bots/19th century that 3-4 approaches/shimaris are bigger than answering 4-4 approaches. I checked Elf and it likes it too (as well as 2-space high, but I'd rather be low). It could also be seen as psychologically successful in disrupting your opponent's presumed plans (which I can empathise with). That's something about strong players, they often are content to play simple good moves, take some territory, give the opponent something because they know such as with your "perfect" black side shape it still has holes to exploit in the future.

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

P.S. when he had facing 3-4 points, didn't Go Seigen like to make the big shimari when one was approached? I don't know if he was unusual in this or plenty of other pros of his time did too, but that's another thing he did that AIs approve of.

Kirby · Post by **Kirby** » Fri Aug 24, 2018 4:20 am

Thanks for the tips, folks.

The specific moves are not as important to me as the underlying point, which I still find to be true: a fair result that I can see clearly is superior to uncertainty. I often aim for refuting my opponent's aims, regardless, and I think a simple, clear result that I can understand is superior.

Your opinion may vary, but I feel this philosophy pretty strongly now.

dfan · Post by **dfan** » Fri Aug 24, 2018 4:30 am

explo wrote:What is KISS?

Keep It Simple, Stupid.

explo · Post by **explo** » Fri Aug 24, 2018 5:38 am

dfan wrote:
explo wrote:What is KISS?
Keep It Simple, Stupid.

Thank you. I looked on sensei's library to see if it was some sort of go evaluation I had never heard of

.

Bill Spight · Post by **Bill Spight** » Fri Aug 24, 2018 5:50 am

Uberdude wrote:Here's what I'd play these days, having embraced the idea from bots/19th century that 3-4 approaches/shimaris are bigger than answering 4-4 approaches. I checked Elf and it likes it too (as well as 2-space high, but I'd rather be low). It could also be seen as psychologically successful in disrupting your opponent's presumed plans (which I can empathise with). That's something about strong players, they often are content to play simple good moves, take some territory, give the opponent something because they know such as with your "perfect" black side shape it still has holes to exploit in the future.

$$B
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . W . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------
Click Here To Show Diagram Code
[go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . W . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . X . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]
P.S. when he had facing 3-4 points, didn't Go Seigen like to make the big shimari when one was approached? I don't know if he was unusual in this or plenty of other pros of his time did too, but that's another thing he did that AIs approve of.

This was my choice, as well.

But then I checked with AlphaGo. Interesting that Elf liked it, as well. Of all the modern top bots, I find that Elf seems to be the one that my intuition is closest to.

That's something about strong players, they often are content to play simple good moves, take some territory, give the opponent something because they know such as with your "perfect" black side shape it still has holes to exploit in the future.

Along those lines, I was amazed at how often Dosaku, giving 3 or 4 stones, would start out building a moyo himself, before venturing into Black's spheres of influence. OC, his opponents never seemed to be able to make much of a dent in his moyos.

Kirby · Post by **Kirby** » Fri Aug 24, 2018 6:27 am

Of all the modern top bots, I find that Elf seems to be the one that my intuition is closest to.

Congratulations

Bill Spight · Post by **Bill Spight** » Fri Aug 24, 2018 8:19 am

Kirby wrote:
Of all the modern top bots, I find that Elf seems to be the one that my intuition is closest to.
Congratulations

It doesn't mean that I am any stronger, but I like Elf's style.

Kirby · Post by **Kirby** » Mon Aug 27, 2018 4:40 am

I decided to play against Leela Zero. I'm not to the level of LZ to play evenly and win yet, and I wanted a chance to win. So I played a handicap game with 6 stones. It was actually pretty easy to win. I'll reduce the handicap next time.

Here is the game:

There is nothing that complicated in this game. LZ invaded the 3-3; I took sente to defend another corner. LZ invaded another 3-3; I took sente to defend another corner. I had influence, and two corners - there wasn't much for LZ to do. I let LZ live where it wanted to live, and made influence.

It's a little uninteresting because of the high handicap, so next time, I'll play with lower handicap.

But for now, it's time to go to work.

dfan · Post by **dfan** » Mon Aug 27, 2018 4:48 am

Leela Zero does not really know how to play handicap games (yet). For one thing, its neural network assumes komi is 7.5. Also, because it thinks in terms of winrate instead of points, it doesn't really know what to do when it thinks its chances of winning are very low (e.g., the beginning of a 6-stone handicap game) and is likely to make crazy moves rather than trying to chip away at the lead.

There has been some interesting work recently to try to deal with these issues, though it is kind of hacky and relies on some implicit properties of the networks. Here's a pointer to the current release; I don't know how easy it is to set up. One important thing is that only certain weight files work well with it.

Kirby · Post by **Kirby** » Mon Aug 27, 2018 5:32 am

Word (am I using that phrase correctly?).

I might give that release a shot, though, from the sound of it I might be better off playing humans for now until research into handicap games has been more fully developed.

Bill Spight · Post by **Bill Spight** » Mon Aug 27, 2018 7:07 am

dfan wrote:Leela Zero does not really know how to play handicap games (yet). For one thing, its neural network assumes komi is 7.5. Also, because it thinks in terms of winrate instead of points, it doesn't really know what to do when it thinks its chances of winning are very low (e.g., the beginning of a 6-stone handicap game) and is likely to make crazy moves rather than trying to chip away at the lead.

There has been some interesting work recently to try to deal with these issues, though it is kind of hacky and relies on some implicit properties of the networks. Here's a pointer to the current release; I don't know how easy it is to set up. One important thing is that only certain weight files work well with it.

I would think that you would want to train the bot against sufficiently weak opponents. Certainly there are previous versions that are six stones weaker than the current best version. But you also have to have some way of letting it know that its opponent is appropriately weak. Winrates assume mistakes. You "have to" assume that Black will make more mistakes than White.

dfan · Post by **dfan** » Mon Aug 27, 2018 8:13 am

Bill Spight wrote:I would think that you would want to train the bot against sufficiently weak opponents. Certainly there are previous versions that are six stones weaker than the current best version. But you also have to have some way of letting it know that its opponent is appropriately weak. Winrates assume mistakes. You "have to" assume that Black will make more mistakes than White.

The way the current LZ experiments work is a two-step process. One is to try to hack in some vague concept of komi, or at least some sort of slider that controls compensation*. The other is to constantly adjust that slider so that the engine is fooled into thinking that it's not in a sure-win or sure-loss situation (so its winrate is between 20% and 80% or something like that). One can think of this slider as enforcing the assumption that Black will make more mistakes than White (putting the slider at "20 komi" kind of means "Black will probably make 20 further points of mistakes over the rest of the game"). When White thinks his chance of winning the game is reasonable, he plays much more reasonable moves.

I don't think there have been many successful experiments of intentionally training against weak opponents, though I don't really know how much it's been tried. My concerns would be that 1) you might learn to take advantage of very specific weaknesses of those particular opponents, 2) you'd have to do a lot of extra training that would take training time away from regular self-play, and 3) either you'd need a separate handicap-play network, or your extra training might interfere with the regular network. On the other hand, humans kind of have a separate handicap-game network (or at least take it into account).

Hopefully this is not too off topic in Kirby's journal!

* Gory but fun details: there's an input feature that says "It's White's turn so the current player gets a 7.5 point bonus", and an input feature that says "It's Black's turn so the current player gets a 7.5 point penalty" (they're separate to make things easier for the network). By interpolating between those values, you can kind of make the network interpolate komi. It doesn't work perfectly, and for some sets of weights it doesn't work at all, but it doesn't need to be that accurate to make the slider idea above work.

Life In 19x19

Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal

Re: Kirby's Study Journal