This guide is oriented towards using Crazy Stone like a chess engine, to analyze and explore games interactively. Unfortunately the UI is actively hostile to this in some ways (much less so as of v1.01), but it's more possible to do than I first thought. Thanks to Cassandra for giving me some tips to get me started.

For some reason, when playing the computer, you only seem to be able to ask for different time controls for you and the computer if you specifically pick a color (rather than nigiri). So if you want to have the computer play at a given rank while you play under a time constraint, you'll have to flip a coin for color on your own and enter it manually.

If you want to analyze your game against the computer, the best time to do it is just after you played (Analyze / Analyze Current Game), since all of the analysis it built up during the game will still be in memory and will be displayed. To load an existing SGF file, use Analyze / Analyze Game Record.

If you do analyze a game you played against the computer, you may want to switch it to Human vs Human mode (via Game / Game Settings) before analyzing interactively. This will let you make moves for both sides and prevent the computer from making moves automatically for you.

You definitely want to bring up both the game record (Analyze / Record Analysis List) and the analysis of current position (Analyze / Move List) and keep them open as you analyze.

The Record Analysis List shows, for every move,

• The move that was played
  
• How much time it took (though it doesn't seem to pick this up from an SGF file)
  
• How many times Crazy Stone considered this position
  
• The fraction of times that Black won when it simulated to the end of the game from this position
  
• A general score estimate
  
• The "width" of the score estimate. The higher this is, the less sure it is about the score.
  
• What move it now thinks should have been played
  
• The difference between the winning probability if its suggested move had been played and the winning probability given the actual move that was just played.

Note that this winning probability is closer to 50% than the "true probability" would be since it is playing somewhat randomly when exploring the game tree, which makes it easier for the side that is behind to catch up.

The Move List shows, for every move Crazy Stone is considering in the current position,

• The move it is considering
  
• How many times it has considered this position
  
• The fraction of times that Black won when it simulated to the end of the game from this position

You generally would always like Crazy Stone to be actively analyzing, which is actually a bit of a challenge since it is easy to kick it out of analysis mode. If the Analyze icon on the right shows a waveform, it's analyzing (and you should see the Move List window update). If it shows a flat line, click on it to start it analyzing again. If it shows the Hint icon instead of the Analyze icon, you got it out of analysis mode and you have to enter it again by choosing Analyze / Analyze Current Game. (v1.01: the Hint and Analyze icons are now separate, so you don't have to worry about exiting analysis mode.)

If you want to look for the biggest mistakes, look at the Delta column in the Record Analysis List and look for the biggest numbers. Go to one move before that move and look at the Move List to see what Crazy Stone thinks better moves might be.

You can explore alternatives but it's slightly painful. First make sure you are in Human vs Human mode so the computer doesn't make moves automatically. Then

• Click on the board window to give it focus
  
• Click on the coordinates where you want to play
  
• "Do you wish to continue the game?" Answer "Yes".
  
• "There is already a move played. Will you delete it?" Answer "Branch".

Now the new move is listed in the move tree on the left of the main window, and the Record Analysis List lists only the moves in the alternate game that would have happened if this move had been played instead.

Unfortunately, doing this takes you out of analysis mode and back into play mode (not anymore in v1.01!). You can do a bit of exploring without going back into analysis mode, because it remembers all of its playouts, so as you make moves, the Move List will actually have information in it, but it will have less and less information as you explore further. To go back into Analysis mode you will have to choose Analyze / Analyze Current Game again. It's pretty annoying to have to go back and forth all the time, but you can sometimes wait a few moves before switching back because it keeps around the information from when it was analyzing the original position.

To return to the real game, go to the move tree in the left of the main window and select a move in the main game from the future. If you select a move in the past, it's still part of your alternative game, so Crazy Stone won't move you back to the main line.

You can print out the Record Analysis List afterward but that's it. Amtiskaw and oren made a couple of tools for processing the output into an SGF file, and I imagine these tools will continue to improve.

As you can tell from the length of this post, it's a bit of a bear working with the interface, but I've found that it's totally worth it. Having a strong "player" obsessively look through your games, and being able to ask it questions about sequences, is super valuable, and I feel like I've learned a ton already just from going over three games carefully so far.

Thank you. I was considering a purchase, but given my lack of time I can spend on the game I postponed it again and again … and now I am happy, because your hints—which sure will be valuable for everybody using CZ—showed me that I definitely cannot, for myself, justify such a purchase at this moment.

And I hope very much that all developers of such software read your post …

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Patience also helps. What you might find is an A B C D situation, where it prefers a play at A, but when you analyze the position with A, it prefers B, and so on, until it settles on a move. So at critical situations in the game you may need to try a few moves and let it sit an analyze for a while. I find this especially true in situations where a group is barely alive in miai, or where there are capturing races with miai. For example, it may spend a lot of time analyzing settling miai and suggest it as a move (even though when it is the opponent's turn the opponent stones never play in the area), so you have to play this miai and then see what the "real" move would be through some of this kind of patience game with the engine.

It has been extremely helpful in analyzing my own games with humans. And playing against it while leaving the "Record Analysis List" open to see what it preferred I play after I play a move has been quite helpful in working on my intuition. The nice thing about this is that although Crazystone has its own ideas you get to see how much worse a move is and often it is very small so you can ignore its suggestion, especially because if the move was viable in a few turns Crazystone will realize its value and adjust the projected score anyway. This makes Crazystone a somewhat style-agnostic teacher, something humans are not very good at in general.

I would like to suppose that CS's strange behaviour (UI; no real "analyse mode") as well as the small extent of analysing tools is due to Unbalance, not to Rémi.

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

That's true !!

Please also note that CS's result (i.e. the favourite moves) may differ from the buffered playouts (got from analysing the previous move) if you let the analysis start from Zero for the current move (e.g. in the main line of the SGF by clicking on any future move in the move window at the left, and then returning again to the current move of interest).

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

Ah, yes, you’re right! I should have written “I hope that all developers and marketers of such software read your post”.

I mean, if Rémi is the “spirit” behind CZ then he should have some influence on how is AI is wrapped up, no? (Or is this about handcuff contracts? But this is going way off-topic.)

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I'm under the impression that Japanese company (like Unbalance) will save some feature so that next year they can release the next version with full price (since stronger AI alone doesn't attract to anyone)
Zen has been doing this for every year since 2010.

Can you do this with Zen? If yes can you make separate tutorial. Thanks.

v1.01 has just been released (use the same download link you got in your original email; the installer will upgrade your version), which thankfully makes analysis significantly less painful. I've updated the post at the top of the thread.

I tried out Oren's script but found that it was too much of an information overload to see suggestions on pretty much every move. I also didn't like the way that CrazyStone reports large deltas on perfectly good moves closer to the end of the game that didn't fit its conception of how to secure the win / engineer a swindle, often much larger than deltas on the bigger mistakes that actually matter in the early and mid game.

To fix the latter problem, I added to Oren's script a fancy transformation on that should make reported win rates correspond much closer to reality, at least for mid kyu and stronger players, stretching out the win rates and deltas near 50% and squashing them a lot as they get outside what previously was 30% to 70%. I browsed through some sample games and as far as I can tell, the new deltas are a clearly better measure than before of the significance and badness of a mistake, and with better deltas it then became possible to apply a threshold on how much of a delta is needed to report the move, fixing the information overload.

Implemented here, in case anyone finds it useful, also submitted as a pull request to the original repo.

https://github.com/lightvector/crazysto ... te-scaling

I actually have Crazy Stone 2012 (lol). Could you explain briefly how turning the engine on and off now works? It is MFOG (where you simply click one button) easy?

Actually it was my script you edited, but yes, it seems like useful improvement.

I'm not familiar with MFOG so I can't make a comparison with it, but the Analyze button is now always up (even if you are in the mode where you enter moves), and if it is on, it stays on while you enter moves.