Page 1 of 3

Crazy Stone Deep Learning first impressions

Posted: Mon May 16, 2016 6:49 pm
by dfan
This is the first Go engine I've used, so I don't know if any of the features I am calling out are standard or are unique to Crazy Stone.

It ran perfectly fine on Windows 7 inside Parallels on my MacBook Pro (quad-core 2.8Ghz i7). Phew!

The UI is a weird combination of slick and clunky (e.g., ugly fonts), but it gets the job done.

I use chess engines all the time to analyze my play, as I don't have a teacher, and they are invaluable for catching my mistakes, both tactical and strategical. I was hoping that the new Crazy Stone would provide a similar function for go. (Of course there are dozens of people here happy to give feedback on games, but I don't want to overuse their good will!) Initial impressions are that it very much can. I fed it a recent game of mine, had it analyze for 10 minutes, and it gave me some important insights I hadn't seen in my own review (mostly having to do with sacrificing stones).

I played a couple of quick games at the 5 kyu level. In both cases I had a comfortable opening lead, got lazy, got tricked tactically in a big life and death situation, and lost. I learned plenty from going over the ensuing analyses, so that's great. I did feel that it didn't play a lot like a 5 kyu human - lots and lots of pushing over and over, very little tenuki. This was just two games though. If I have to play people (or Crazy Stone on a higher level) to get more interesting fuseki, that's okay. On the other hand, in one game it "misread" a relatively straightforward life and death issue in a human sort of way, and so did I; in the analysis it was happy to point out what it "missed". (Scare quotes are all because of course it would have gotten it right running at full strength.)

The process of analysis was slightly awkward. I might have missed something in the UI, but I couldn't seem to get it into a mode where I could just make moves on the board and have the computer's opinion constantly update. If I made a branch of the game tree, I would be put back in game-playing mode, where I could ask for a hint but not see the engine's current analysis. Once I made a move, I could see what it thought of it, though.

Anyway, I'm quite happy with it so far. I think that having a strong player always around to point out my worst mistakes within minutes is really going to help tighten the feedback loop of learning. Of course, at my level, I'm not in a position to assess its true strength.

Re: Crazy Stone Deep Learning first impressions

Posted: Mon May 16, 2016 8:05 pm
by LokBuddha
I'm a bit rusty but I am 6d almost 7d on wbaduk. I played poorly in this game, no reading at all, and miraculously lucky win?

Enjoy the blood bath.

Re: Crazy Stone Deep Learning first impressions

Posted: Mon May 16, 2016 11:51 pm
by Cassandra
dfan wrote:The process of analysis was slightly awkward. I might have missed something in the UI, but I couldn't seem to get it into a mode where I could just make moves on the board and have the computer's opinion constantly update. If I made a branch of the game tree, I would be put back in game-playing mode, where I could ask for a hint but not see the engine's current analysis. Once I made a move, I could see what it thought of it, though.

Yes, indeed, the process(es) of analysis are a bit awkward. And you did not miss anything in the UI.

I would like to suppose that Unbalance still put their main emphasis on "playing", and not on "analysing". Hopefully this will change in the future, when it might become easier to add some additional features for analysis than to add decisively to the playing strength of the programm.

+ + + + + + + + + +

But let's come back to your experience.

"Analyse Current Game" is as long in a "continuous mode", as you browse through already EXISTING moves in the SGF. It does not matter whether you are in the main line, or change to a sub-variation that is availably through the move window at the lower left. CS still goes on in "analysis mode".

Please note that the results of the playouts are buffered and used in / transfered to the analysis of the next move. This might lead to a "Move List", where (the next) move #1 (in a series of moves) already has about 100.000 playouts, but all other moves from #2 on are in a range about several 100 playouts.

If you want to get a better idea of which alternative moves CS had "thought about", simply go back to the last move before your move of interest. Analysis will start from Zero again, and usually you will see several moves with a noteworthy amount of playouts at the top of the "Move List".

* * * * *

However, once you wanted to try out a variation (= added a NEW move to the SGF), CS seems to go back to "playing mode".

In this case, if you want to further study a variation, with following CS's suggestions for its "favourite moves", you might first click the moves that are shown as #1 in the "Move List" (after having started "Analyse Current Game", and waited some time until you reach a suitable amount of playouts for the first #1 shown), as long as the number of playouts shown makes sense (it will decrease continuously).
I think that this will show you something similar than CS's supposed main line for this sub-variation.

When you then go back to the sub-variation's root thereafter, you can start "Analyse Current Game" in the menue again, and you will stay in the "continuous mode" for analysis, as long as you follow an already known path.

This means that you have to do it every single step by single step (= starting "Analyse Current Game" again and again), if you want to base your next NEW move on CS's current evaluation.

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 9:47 am
by oren
Has anyone seen a way to save a game analysis? I've only had a little time to play around, but I haven't seen anything obvious. Ideally for me would be something in the SGF comment field with percentages and "CS Best". I'd be ok doing the conversion if there was at least some way to save the analysis as a text file.

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 9:57 am
by Cassandra
oren wrote:Has anyone seen a way to save a game analysis? I've only had a little time to play around, but I haven't seen anything obvious. Ideally for me would be something in the SGF comment field with percentages and "CS Best". I'd be ok doing the conversion if there was at least some way to save the analysis as a text file.

Probably the only way currently:

-- Several screenshots of the list
-- OCR

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 9:57 am
by oren
Cassandra wrote:-- Several screenshots of the list
-- OCR


That was my fear. :)

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 10:22 am
by sparky314
Would be cool if they could add a heat map feature for analysis mode.

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 10:32 am
by palusijf
File -> Print -> Game Record Analyse

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 10:42 am
by Cassandra
palusijf wrote:File -> Print -> Game Record Analyse

Thank you very much indeed.

Never used "File" in the menu so far ...

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 10:49 am
by palusijf
You can then print with pdfcreator or on a printer.

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 10:59 am
by oren
Thanks! I actually never considered print before. I can probably make this work for me.

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 11:21 am
by Cassandra
oren wrote:Thanks! I actually never considered print before. I can probably make this work for me.

The following worked for me:

-- "Mark all" in the PDF
-- STRG + C
-- STRG + V in WORD (I suppose any other similar text processing programme will do, too)
-- Convert text to table (separator = " ")

The "columns" in the PDF are separated by single space characters.

However, please note that there also is a space before single digit numeric coordinates.
Their format is e.g. "N,11" or "B, 2".
Therefore, before transformation, you will have to replace all ", " (comma & space) by "," (comma).

Re: Crazy Stone Deep Learning first impressions

Posted: Tue May 17, 2016 12:51 pm
by sorin
LokBuddha wrote:I'm a bit rusty but I am 6d almost 7d on wbaduk. I played poorly in this game, no reading at all, and miraculously lucky win?

Enjoy the blood bath.


I wonder if those 9 stones handicap played a role :-)

Re: Crazy Stone Deep Learning first impressions

Posted: Wed May 18, 2016 12:38 pm
by oren
I made a script that took the recording analysis from pdf printout and the original sgf to make comments and put CS's best move in as a triangle. The deltas aren't quite as big a swing as some of my bad kyu games, but I thought it was still fun to show off.


Re: Crazy Stone Deep Learning first impressions

Posted: Wed May 18, 2016 1:08 pm
by erislover
oren wrote:I made a script...
Can you share it?