Hi,

I am glad to announce that my web site at https://www.crazy-sensei.com/ is now able to analyze sgf game records and produce an analysis that can be browsed online. In order to test it, I will analyze some games of this forum and post the links.

A little bit of explanation of the analysis:
- The blue line is Black's probability of winning
- Red bars indicate errors and their magnitude
- You can click on move numbers on the board to browse the search tree

Rémi

Nice tool! I signed up and tried to analyse the last game from my study journal at 6d level (10 credits) but it said I didn't have enough credit despite my account showing 10 credits, an off-by-1 or > vs >= error? 4d level worked and is in progress so let's see what that says, but seeing as the players were 4d I wanted to try the deeper analysis.

Also as a testing point that sgf has variations but the game is the mainline. Will those get analysed too or confuse it or get ignored?

P.S. funny that for the first move the policy network strongly prefers the large knight's move away from tengen over other nearby moves, as well as tengen itself and normal corner moves.

You have 9 credits if you posted one 4d game. Maybe you navigated back to your profile. If you reload, you should see 9, because the 4d analysis costs1. You have 10 more now. I also added 10 credits to everyone who registered so far. I may add more later when my server is not busy any more.

Note that a 6d analysis may take time, especially when there is a queue. You can set an option in your account profile to get an email notification when it is ready.

I tried the 6d level analysis first and it failed, and went back to check my profile and it still said I had 10. I then did the 4d one and it worked and credits went down to 9.


Thanks!

Are you planning to allow the purchase of credits with real money in future? I wonder what the exchange rate would be...

Btw, I really like the policy view, great way to visualise where it wants to play and how much with the different sizes.

That was a bug. Thanks for reporting. The fix will soon be online.

Yes.

It seems that some people register their email address, but never set a password. If you don't receive the registration email, then please tell me by sending a mail to the contact address, or by telling me your provider here. I am still trying to configure my mail server so that it passes the various spam filters.

Also some people obviously made typos. I have one @gmail.con address, for instance. Trying to register again with the correct spelling should work.

And if you wish to change your password: log out, then choose "forgot password". Then you'll be able to reset it.

Thanks

Thks Rémi,

I confirm that you're allotted 0 credit at first.
Only after uploading a game with 4d analysis is the store credit filled, decreased by 1... and showing 9 then .

I also have 0 credit but I'm not in a hurry

So 4 dan use pure neural net and 6 dan toped with a little bit of MCTS playout?

Rémi, does this tool read the komi from the uploaded SGF and take that into account? I noticed on my game (#34 and #43) (in which I left komi blank as 0) it thought black was leading a bit from the start. I recall reading that some bots like CrazyStone and AlphaGo slightly prefer white with normal komi (6.5 or 7.5?).

Yes, it reads the komi from the SGF.

It did not work for me. When I tried to upload a game, it always says that I have no credit. Have I missed anything?

Looks cool. As I am waiting for one of my games to be analyzed, I have a feature request: give some sort of feedback when it is reasonable to expect that the game will be processed.

_________________
Occupy Babel!

Thanks for your suggestion. I'll think about providing an estimation. You can set an option in your profile to get email notifications of when the analysis is ready. Your game should be available in about 15 minutes.

I turned that setting on and already forgot I'd done so! I'm just the sort of annoying user you want testing things, I guess.

_________________
Occupy Babel!

Thanks; this is great! I've thought about purchasing Crazy Stone in the past, but all I really want it for is analysis. If this works well and the credits are priced reasonably, this is definitely something I'll use.

Do you want us to post feedback about our experience with the tool here or elsewhere? I'd like to be helpful in any way I can!

Is there any chance that some of these analysis tools (like the Policy view and the red 'error' bars) might be added to the PC version of Crazy Stone Deep Learning? The visuals make reviewing a lot easier, and it would be a shame if the $80 pricetag for CS:DL is for something that is worse for game review than this online service.

Hi Rémi, I'm really liking this tool, a nice UI that works well on mobile too . I have a question about the meaning of the policy view red squares. Is it how popular that move is in the policy network, irrespective of winning probability after playing it? If you look in this game at move 15, for the next move white (Haylee) played the 2-4 point in the corner which is a very normal move in pro games, and it looks like it has the biggest policy red square. However, that move has a medium-size red error bar, indicating it was a mistake (according to CS) and raised black's winning chances. On the 'tree' view it is #2, and #1 is the 2-space pincer (which has significantly smaller red policy square). Does that mean the pincer was the best move for win probability, with the corner move 2nd? This does actually tally with some other bot analysis I heard about (maybe AlphaGo, or Zen or CS, I can't recall) that it didn't like that corner defence.

I am afraid that the awareness about the value of analysing tools is greater in the West (e.g. Rémi) than in the East (e.g. Unbalance).

In my opinion, it should not be underestimated that the analysing tools of a programme (e.g. CS DL) will also benefit players of a strength (far) above the programme's playing strength.

+ + + + + + + + + + +

I let Crazy Sensei analyse one of the recent AlphaGo games (vs. Meng Tailing, #9). It seems to me that the result (winning probability) is totally congruent with CS DL's "P B(wins)".

But with Crazy Sensei, you have additional views, and the histograms are also stored for browsing through.

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

Whenever you have time to do a few UI/UX tweaks:
- There is no way to know the current move number.
- Keyboard shortcuts (at least left/right to go over the game) would be nice.

My current workaround:
- right-click on current vertical bar in the win% graph
- Inspect element (in dev tools / firebug etc)
- find value of 'data-title' attribute