What happened is that you had 2 instances of Leela running at the same time on your computer (so they where not at the same game move). One that was performing the analysis, and one that was playing as white (with the second in advance because it does not have to analyse black moves). In this case, yes they won't necessary agree on the best move.

If you want to have only one instance, you needs to select the bot that is labelled "Bot used for the analysis (XXXXX)":


The drawback in this case is that you have to wait for the bot to finish the analysis of its move AND your move before it continues playing So it's fine only if you take much more time thinking about your move than it does. (<- which is the case for me )

Note that if more than one bot has to be used, then you will see the option to make sure they don't think at the same time appears (see, the check box "No overlap thinking time" is not on the first image above):


At the beginning, I had considered something like you say: play move 12, then go back to move 3 for the analysis, then move again to move 14 if the opponent played move 13. But yes, it's a big big headache.

In fact, I already had a very very hard time implementing the "Undo" feature considering the analyser and the players "work" on the same SGF tree (players extend the first branch, analyser adds other branches at each node, so one need to be careful when removing nodes of the first branch).

The way I am using it now is having the slow instance performing the analysis, and the fast instance playing white or black. I don't use the thinking time limit any more, only the playouts limit from the command line. Plus, on Linux (but should be doable on windows) I set the analyser instance as a low priority process, so it won't slow down the fast instance. But because I use a playout limit without time limit, the quality of the analysis is still the same.

Yes, Human VS Human is what I imagine as "spectator mode" or "kibitz mode"

There are still plenty of work in progress before the next release, I have plenty of things to polish, and I will have to spend time on the documentation as well.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

Oh, I see now, thanks for clarification! I will use "Bot used for the analysis (XXXXX)" when necessary.
Me too I also like using constant number of playouts/visits instead of time settings just because of these overlapings. Or maybe something else you're doing on your computer.

I spent quite a bit of time investigating this issue with the help of snippy.

As it appears, the issue was due to Leela value network unable to estimate games with too high number of handicap stones. In this specific case, the value network was giving black winner by a lot, and the Monte Carlo simulations were giving white winner by a lot. The winning rate (a function of both previous estimations) was around 50%, slightly above.

This pushed me to implement graphs and "delta graphs" (<- this is the new name I gave to the "Comparison graph for black/white") for Value Network and Monte Carlo estimations. In case of doubt with the win rate, it is better to double check the two other "metrics":

Ray, AQ and Leela have the three metrics available. Leela Zero has only the Value Network one (that is wrongly named Win Rate at the moment). (GnuGo has none of them )

As I was investigating this issue, I found an unexpected result with AQ and Ray. Both of them, in Chinese rule set, in handicap games:

• Do not add the number of handicap stones to white score (this could be ok depending of the rule set used)
• Only consider a komi of 0.5, even if the komi was set at a different value

Leela, LeelaZero and GnuGo all account for the number of handicap stones when calculating white score.

I opened issues on both projects [1][2]. I mention that so that until it is fixed, be careful when playing handicap games against Ray and AQ.

I also noticed some behaviour that may be a violation of the GTP protocol for Leela and Leela Zero:
When they choose to resign, they create a move in their move history that requires to be undo if one want to go back in the game tree.

• GnuGo does not consider "Resign" as a valid move.
• GoGui also agrees with GnuGo interpretation of the standard.
• Sabaki avoid this issue by replacing "undo" by "clearboard" and replaying the whole game up to the previous move.
• Ray does not undo.
• AQ undo command is broken at the moment (did not check for the previous version behaviour).

I think this is an area where the GTP protocol documentation is not clear enough. I did not opened an issue for this one, seems everybody has its work around. But clearly, this poses problem when using Leela/LeelaZero with Gogui.

This was giving me headaches for the live analysis with Leela/LeelaZero. This is for such cases that I am happy I kept GnuGo around. It's implementation of the GTP protocol is somehow the reference implementation. So when it works with GnuGo but not for another bot, I can start to look into such details.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

Apparently, this has been fixed for Leela Zero in v0.10, v.11 or v.12

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

This is me again!

A gentleman on Github has proposed a different layout for GRP, that integrate the graph inside the main view:


He went as far as to make a modification of the code to get a better idea. It would looks more or less like that (consider this as a draft, it would need some more polishing):


I also believe GRP GUI would require a new layout. I would like to ask all of you for suggestion, comments, and so on. Maybe I should start a different thread for that, but I don't want to spam L19 with GRP threads

For reference, the initial discussion is on github there.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

I like this new layout idea. But only wonder how would graph fit in such a small area, would you be able to see red and green deltas?

Does anyone know how to use updated versions of leelazero with GoReviewPartner? (current weights and/or .exe)
Edit: I think I've figured out the weights, but not how to use updated leela versions.

It works the same way as for example, Sabaki:

• Download and unzip Leela Zero
• Download the weight file you plan to use
• In GRP setting panel, Leela Zero section, indicate the path to leelaz.exe and the parameters tu run it in GTP with proper weight file.

Beware that latest version of Leela will forcefully run the tuner when starting the bot the first time. This will block GRP for a while.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

Always I get an error saying
ERROR: Leela Zero did not identified itself as expected 'Leela Zero' != ''
and nothing happens

But when I completely replaced the Leela Zero folder with a fresh download of the windows 64 (notcpuonly) it is working. I'm not sure if the type of leela I downloaded mattered (earlier I downloaded cpuonly as I wasn't sure if my computer had a gpu or if there was something else I did differently. Anyway it all seems well now.

By the way, thanks for the help and amazing program!

The error message above typically appears when Leela Zero fails to start in GTP mode. It could be that the program fails because it cannot find the weight file, or the --gtp parameter was not added.

On the next version, I have added a handful of error detection for Leela Zero, related to weight file, this will hopefully avoid more headaches.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

What is this? Installed Python, use absolute path for Leela. Weightsfile is "bestnet" here. Saving the settings in UI it says below. It's on Win7x64, thanks!

Path: c:\Users\Michel\programme\Drago\Engines\LeelaZero\leelaz.exe
Settings: gtp --noponder --weights bestnet

Hi!

This is normal: originally, GRP used the tkinter (GUI tollkit) file pickers (the widgets used to select files to open or save). Then a contributor sent a patch to replace those tkinter widget by the file pickers of wxpython, another GUI toolkit, because the latter are more beautiful and easier to use. They are the same as the system, with your usual bookmark. So if you happen to have wxpython installed on your system (which is often the case for Linux) then it will be used. Otherwise, GRP uses the usual tkinter widget.

In fact, for the future, I am considering moving entirely GRP to wxpython.


Yep, that was a bug in the development version.

I just published the version 0.11 of GRP, you can download it for windows from here: http://ns1297.imingo.net/ (make sure to refresh the page, F5, in case you visited it recently). I haven't finished updating the documentation, so no announcement yet.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

Hi, is it possible to provide an option to download GRP with CPU-only Leelazero? Thank you!

Hi there,

would it be possible to change the table behavior in such a way, that it auto-updates when you step through the game moves? For me it will stay on the move that it has been opened with!

Take care and thanks a lot!

Hi, I'm using the Leela Zero version from http://ns1297.imingo.net/

The below text boxes do not scale when the window is stretched vertically. Text boxes end with "Black Value Network win probability:" - for the interesting part I have to scroll. Instead, GRP shows plenty of unused space. Can this be changed?

I used "Convert RSGF" on a rsgf file it just created, this is what happens:

Hello:

I just want to say THANK YOU to our friend pnprog, for all the work done. Excellent!

Hi!
I was out of town last week, so no time to work on GRP.


Yep no problem. I think the 64bit version is fine?

I will see what can be done. Those two text boxes are really not satisfying

The exact same problem was reported on Github, I will investigate this.

Yes, I will change that.

Thanks you! Encouragements are much appreciated

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

Thank you! 64 bit is fine

Hi everybody!

So I finally found the time to complete the documentation of v0.11, so here is the official announcement

The big change for that version are:

• each bot can be configured in 2 profiles, a slow profile (good for analysis) and a fast profile (good for real time play)
• the new Live Analysis mode

For the live analysis, I am going to copy paste the description I put in the documentation:

A live analysis makes it possible to record a game or play against a bot, while, in the background, having the analysis proceed. This is a way to save on the analysis time, so that when the game is over, part of the analysis is already completed, and the review can start immediately.

It also has the advantage to make the result of the analysis directly available (given the bot performing the analysis is fast enough to catch up with the game speed) which is interesting during a live review.

So first, you have to select the bot that will perform the analysis. It is probably better to select a slow profile bot.
Second, you select who will be playing as black and white. You basically have three options:

• Human : you will enter the move one by one for that color.
• Bot used for the analysis : you want to reuse the bot performing the analysis in the background to play for that color as well
• Another bot : that other bot will play for that color

So here are different way to use the live analysis mode:

• Human VS Human:
  
  • You can use that mode when watching a game online or at a tournament. If the bot used for analysis is faster than the players, then it will provide you with real time data about the on going game, like win rate, and different possibilities for the on going moves.
  • You can use it when playing across the board, to record the game. At the end of the game, then the analysis result can be used to discuss the game with your opponent and check out your respective mistakes.
  • It can also be used as well when reading a game commentary: you enter the moves as you read the commentary, and play the variations in the open position mode. Then you can stop at different moves and check the bot own opinion against the commenter opinion.
  • Of course, the bot opinion is relevant only if the bot is stronger than the human players.
• Human VS Bot: This is mainly for training. You play the bot, and at the end of the game, you review the analysis. You can also stop mid game and check the analysis of the first part of the game.
  
  • You can have the bot playing with you different from the bot used for analysis. Then it makes sense to have this bot weaker than the bot performing the analysis:
    
    • You can use the slow profile version of a bot for analysis, and the fast profile version of that same bot to play against you.
    • If your level is DDK (11 kyu or weaker) then you can play against GnuGo, and have Leela to perform the analysis, and point out your mistakes and how to avoid them, but also GnuGo's mistakes and how to exploit them.
    • You could also set Leela Zero with an old weights file as her fast profile and using the latest weights file for her slow profile for the analysis.
  • You can have the bot playing with you to be the same as the bot used for analysis. This is interesting to save on computing, because the bot analysis is performed at the same time it is choosing its next move. But in that case, you will have to wait for the bot to finish analysing your last move before it starts thinking about its next move. Also, because it plays and analyse at the same time, the Delta for its moves will all be at zero (meaning it considers its move the best moves).
• Bot VS Bot: this is a possibility, but I am not sure for what use case...

When a bot is selected as black or white, that is not the same as the bot used for analysis, or when black and white are played by different bots, then an additional option appears No overlap thinking time that indicates to GRP to make sure the bots are no thinking at the same time, so that your computer processing power (CPU, GPU, RAM) does not becomes a shared resource.

Depending of the situation, this option will be or won't be useful:

• If the bots have a limited number of play-outs, then it is not really necessary
• If one bot run on CPU and the other one on GPU, then it is not that necessary
• If one of the bots is ran across the network from another computer, then it is not really necessary
• If the bot has a limited thinking time, then you should use it

Generally speaking, it is best to use play-outs limits for your bots. You can also set the process of the bot performing the analysis to run at low priority, so that it won't make your opponent slower.



During the live analysis, there is an Open position button available at any time to try out variations. And when the analysis for the 2 first moves is completed, a Start review button will appear that starts the usual review panel for this game. Of course, only the moves already analysed will be available, but the review panel will update itself automatically as the analysis progresses.

So basically I have been using this mode a lot since i started implementing it, playing Live against an older version of Leela Zero, and having our game reviewed by Leela. To me this was eye opening.

Some of the other improvements since v0.10:

• More graphs available
• Quick evaluation in Open Position mode
• Auto update of the Review Panel and graph to follow the analysis
• All analysis data are now saved as SGF (non standard) properties
• RSGF to SGF converter (at the moment now really working apparently)
• A GTP tester (like a terminal) to quickly check the bot is running fine and find out issue, because at the moment, issue with bot settings is a big obstacle to the adoption of Leela Zero (and other bots)
• More check for Leela Zero to handle configuration issues (opencl tunner, wrong weights files...)
• A table to summarize the data off the position and all variations
• Several options to display the variations labels and colors
• Fixed plenty of bugs
• Added plenty of new bugs

In fact, I now realize I tried to put too much things into this version, and by doing so, I kept delaying and delaying and delaying this new v0.11, which is detrimental to GRP. More frequent small versions are better to try out the features and find bugs. In fact, there are more bugs than usual in this version, because it's hard for me now to try out all configurations options and combination of those options.

So I will probably release a v0.11.1 soon after I fix all current issues, before going on implement more features.

For the next version v0.12, the main feature I plan to add is support for generic GTP bots that can be used in Live Analysis (and Open Position). Now, we have plenty of very strong bots to analyse our games, and it's good also to have some weaker bots to play against (while leaving the analysis to the strong bots of course).

Then Ideally, that v0.12 will be the last version before v1.0
For v1.0, v1.1, v1.2, v1.3... I would like to stop adding new features, only maintenance:

• Adding support for new strong bots when they appears
• Bug fix
• Translation

The last point is important to me: I received a translation in German after v.10, but with all the interface changes I made since then, that translation is all broken. I feel bad about this, and I cannot really fix that by myself. So I would like to stabilize the programme (at least the interface) so that any work on translation won't be broken in the future, before asking for help for translation.

Then, once V1.0 will be out and living it's own life ("in production" as they say), I will probably slowly start a v2, that will be a restart from scratch, avoiding the design issues I made with v1, probably using wxpython instead of tkinter, python3 instead of python2, and that one will be opened to new features.

So, here you are:

• Documentation (make sure to refresh the page)
• Code source
• Windows download (Leela Zero no GPU coming soone)

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

I just added the link on the download page.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!