Life In 19x19

in another thred, remi coulom kindly proposed to analyse some game with crazystone (CS), and I proposed my latest rout.
You can find the result here

i find the result extremely interesting at this level (6k): tactical blunders clearly appears as move with a huge delta (delta is the difference of winning proba between crasystone best move and the game move)

and the territory estimate is also interesting;

I was surprised to see many good shape move proposed by crazystone.
I thought that computers cared little about shape and just made things works; Also sometimes i felt that CS respected strategic principle much better than the puny (weak)humans: See for example the attachment against the corner proposed at move 16 that seems to use "attach to your opponent to defend"

I find it fascinating as an emerging property of just random playouts

Also, CS suggestions during fuseki seemed really reasonnable to me compared to for example the game with fuego on this forum where the computer played very strange moves;

Overall i am impressed, by the human-like quality of the proposed moves (i am not really qualified to be impressed by their accuracy ie i cant judge if suggestions are good) but still there are moves i don't understand, so here is the game annotated with CS suggestion and my comments / questions;
There was a huge ko around move 130 with a few weird things in CS suggestion, too

I would be extremely interested to see if strong players agree with CS analysis or if they find it, well, crazy.
maybe now CS can beat you but you can stil analyse better than it ?

I am really looking forward to compare the human guts and the silicon playouts opinions on a weak game like that


(i did not report ALL CS suggestions, follow the link above if you want everything)

I was BLACK (and not proud of this game)

I was also impressed by Crazy Stone's analysis of my game I posted in http://www.lifein19x19.com/forum/viewto ... f=4&t=5277 . I'm not convinced it has given better moves around move 42 we were discussing, but it did find some mistakes later in the lower right corner fighting but I've not had time to analyse and write it up yet.

My opinion on the start of your game:
- I'm not sure f5 is better than f2.
- q6 and e2 both good
- yes should f5
- CS q5 seems early?
- m3 indeed a mistake, inefficient with k4, CS's q5 is good.
- CS is correct q4 is a skilful sabaki, all its suggestions here are good.
- Your r5 is really bad and helps white, r6 is shape. This kind of thing just screams bad style to a human, but bots do seem to like this sort of bad style sente too. EDIT: just checked full anaylsis: it did suggest r6 .

This does indeed look like an excellent analysis tool. Of course it can't explain the why like a human, but the suggestion and expected sequences certainly give a good shot at it.

In your commentary to move 29, you reached the following position:

Click Here To Show Diagram Code

[go]$$Bc
$$+-------------------------------------+
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . O . . . . . , . . . . . X . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . , . . . . . , . . . . . , . . .|
$$|. . . . . . . . . . . . . . . . X . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . X X O . . .|
$$|. . . O . . . . . . . . O X O X O O .|
$$|. . . . . . . . . . . . X O O X X O .|
$$|. . . O . . . . . X . X 1 . . X O . .|
$$|. . . . O X . . . . . X O O . X O . .|
$$|. . . . O X . . . . . . . . . X X O .|
$$|. . . . . . . . . . . . . . . . . . .|
$$+-------------------------------------+[/go]

Why not
instead?

Unfortunately, I'm not strong enough to analyze the rest of the game.

Fedya wrote:In your commentary to move 29, you reached the following position:


Why not

$$Bc
$$+-------------------------------------+
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . O . . . . . , . . . . . X . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . , . . . . . , . . . . . , . . .|
$$|. . . . . . . . . . . . . . . . X . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . X X O . . .|
$$|. . . O . . . . . . . . O X O X O O .|
$$|. . . . . . . . . . . . X O O X X O .|
$$|. . . O . . . . . X . X . 1 . X O . .|
$$|. . . . O X . . . . . X O O . X O . .|
$$|. . . . O X . . . . . . . . . X X O .|
$$|. . . . . . . . . . . . . . . . . . .|
$$+-------------------------------------+

Click Here To Show Diagram Code

[go]$$Bc
$$+-------------------------------------+
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . O . . . . . , . . . . . X . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . , . . . . . , . . . . . , . . .|
$$|. . . . . . . . . . . . . . . . X . .|
$$|. . . . . . . . . . . . . . . . . . .|
$$|. . . . . . . . . . . . . X X O . . .|
$$|. . . O . . . . . . . . O X O X O O .|
$$|. . . . . . . . . . . . X O O X X O .|
$$|. . . O . . . . . X . X . 1 . X O . .|
$$|. . . . O X . . . . . X O O . X O . .|
$$|. . . . O X . . . . . . . . . X X O .|
$$|. . . . . . . . . . . . . . . . . . .|
$$+-------------------------------------+[/go]

instead?

Unfortunately, I'm not strong enough to analyze the rest of the game.

Yes you are right of course, one more reason to go for the cut for B

@uberdude thanks for the comments, very useful as always

perceval wrote: I was surprised to see many good shape move proposed by crazystone.
I thought that computers cared little about shape and just made things works; Also sometimes i felt that CS respected strategic principle much better than the puny (weak)humans:

Perhaps misunderstanding what was going on? Looking at in the wrong way because of how we humans think?

Try from this perspective. Why do we humans tend to make the move we have learned is the proper shape? Why do we humans respect the strategic principles we have learned? Because when we play that way (as opposed to violating these) we win more games. It's the direction of our learning.

The program is working this in reverse. Asking "from which move do I win the most games" (by trying lots of semi-random playpouts between two equal players). So would you expect the answer it came up with to often violate "proper shape" or "proper strategic direction"?

Where these programs surprise us with "strange" moves is where the total distributed aji of a move makes it the best move. It's one thing to notice "that move would be a ladder breaker" (for one ladder) but quite another where the benefits are in many small pieces all over the board.

BTW -- it is not impossible to have one of these programs able to give us a "why" in human understandable terms. Maybe not for every move but for most of them. Remember, before MCTS supplanted them, rule based AIs had reached about 6 kyu. While far inferior in choosing which next move is best (from among all the moves that have some obvious go reason behind them) if given a particular move in that set, can tell you why that move was selected to be the set. Just because a 6 kyu couldn't imagine the next move selected by a pro, if watching the game, most of the time upon seeing the move can say "wow", that's a great move" (and say why). Harder are the situations where a trade is involved where it can look like too much being given up for too little (but where that "little" plus sente plus the aji remaining makes that the best move).

So even the more limited analysis Crazy Stone can provide is useful. Most of the time, shown the "better move" will be able to immediately see why better.

Mike Novack wrote:Try from this perspective. Why do we humans tend to make the move we have learned is the proper shape? Why do we humans respect the strategic principles we have learned? Because when we play that way (as opposed to violating these) we win more games. It's the direction of our learning.

Not quite. Whether particular shape moves result in more wins isn't the kind of thing we can track--it's not something we can observe in any given way. Rather, we observe that the shape move gives us a certain (usually local) benefit by preventing the opponent from playing other moves, giving us liberties, etc, etc.

But that's not really the crux of what you're saying.

hyperpape wrote:...... Not quite. Whether particular shape moves result in more wins isn't the kind of thing we can track--it's not something we can observe in any given way. Rather, we observe that the shape move gives us a certain (usually local) benefit by preventing the opponent from playing other moves, giving us liberties, etc, etc.......

I didn't mean in that way (some particular shape).

I meant we have learned that when we pay attention to shape, pay attention to the strategic direction of the game, etc. then we win. Or that being less certain, have learned that when we neglect those things we will lose.

Even with the qualification, I think I disagree. I guess it's true that it was eye-opening when I saw how I could beat my first real opponent by playing tenuki. But I only got to that point because I read a book that talked about the subject and it made sense and I could see why it worked.

Are one able to do get such a analyze from the program or is it just the programmer who where able to draw out some data, through a back door or something?
Just to put in a sgf file and get such a review would become handy in many cases. Might be a risk that one take the review for an "answer" but in late endgame it must be quite accurate.
(I have heard about the OS called Windows but I'm not in love with it. Maybe it is time to make an exception....)

remy coulom (crazy stone author) offered to take some sgf and make the analysis, and published a web page with the results. maybe the functionality will be on the commercial version later but not now if i understood correctly (but i do not own the software myself so not sure)