Announcing GoReviewPartner - v0.15 with Pachi and PhoenixGo

[djhbrown]

yup - that was it. thanks.

[pnprog]

Hi everybody, that's me again!

One more update regarding v0.6 that should be released this week. Leela 0.10 has been released last week, and GoReviewPartner v0.5 works just fine with that latest Leela. The new Leela GUI now displays the moves sequences with numerated stones on the Goban just like uses to do GoReviewPartner. I would like to believe that GoReviewPartner prompted Leela's author to implement the same feature, but that's probably not the case

I was reading the comments about Leela on Reddit, and somebody mentioned an improved influence GTP function. I had not noticed that Leela had such function, so I just had a look and implemented the result into GoReviewPartner the same way GnuGo territory is implemented. Here is the result for the same position as in the previous screen-shot for GnuGo:

Screenshot0001.png (52.8 KiB) Viewed 20614 times

Two things to notice:

• It's an influence function, not a territory function (I will rename the button) so comparing GnuGo and Leela maps is probably not a good idea
• I noticed Leela provides the data instantly when asked for the map, it must be some sort of fast heuristic (so no simulation or whatever)

As for what I mentioned in the previous post, about completely separating GnuGo and Leela, this is now completed. So before running the analysis, the user is asked what bot to used:

Screenshot0002.png (21.53 KiB) Viewed 20614 times

This forced me to make some big code factorization, but the resulting code is (hopefully) much cleaner and it should be easier to finally implement support for additional bots.

All this new code has been pushed to GitHub already, but not heavily tested yet.

[djhbrown]

such a fickle friend of the gnu - you have now made it possible for the discerning customer to ignore the old cow completely.

at last!... thank you very much!

[johnsmith]

I also added the neural network value. But I am not sure what it really means

Can you please add "Net Win %" instead of "Net Prob %" (it's the last row on the right screen)? It's very useful number to evaluate the game, much better than MC... Can you please add it on the left screen too (for the current position as well as variations)?

Thanks for contributing to the go community with this app! I find it very useful and use it very often to analyze what was the best sequence in my tournament and online games. I tried replacing Leela 0.9.0 files with 0.10.0 and it works just perfectly.

[goame]

@pnprog thx it looks great.

Can you create an ios version?
It would be great to have it on the ipad pro 12.9 screen and especially on the ipad pro 2 and iphone 8.

[pnprog]

Hi johnsmith,

Thanks for your positive feedback, it's very motivating for me!

Can you please add "Net Win %" instead of "Net Prob %" (it's the last row on the right screen)?

If I am not wrong, it is what GoReviewPartner calls black/white win probability, or am I wrong?

[pnprog]

Hi goame,

@pnprog thx it looks great.

Can you create an ios version?
It would be great to have it on the ipad pro 12.9 screen and especially on the ipad pro 2 and iphone 8.

It is unlikely that Leela will run on ios anytime soon (if ever). So the analysis part would not be available on ios or android.

It would be possible to have a application of ios/android for reviewing the *.rsgf files just like a viewer (without possibility to test alternative sequence against Leela), but it would have to be another project, because the graphical framework I use (tkinter) is not available on ios or android.

One possible option would be using Kivy, another graphical framework dedicated to touch interface. I did experiment with it and made some small go applications on my android phone last year, so it would be possible. But the "work-flow" would be awkward: running the analysis from computer, then transferring it on the tablet, then reviewing it. Plus, some "mouse gestures" (like "mouse over" or "mouse scrolling") would not be available on tablet, so the interface would have to be rethink from scratch.

So... yes, another project

[pnprog]

Hi everybody!

Here come v0.6 !

So v0.6 does not really bring new features. In fact, this version is more about "polishing" the software (code cleaning, code re-factorization, better error handling) and is a big step forward v0.7 that will add support for Alphago that will add support for Ray.

So what's new:

• Separate analysis for Leela and Gnugo
• Territory map for GnuGo and influence map for Leela
• A few more settings to review a game
• Possibility to export an image for the gobans (postscript for now, i will try to change it for png later)
• Tons of bug fixes
• v0.6 for windows ships with Leela 0.10 (but v0.5 was working fine already)

Same as before, there is a direct download for GoReviewPartner+Leela for Windows users.
For Linux/Mac, it is best to get the source code from Github.

So next step is adding support for Ray, zakki, the author of Ray, has just added a feature (a GTP command) to allow GoReviewPartner to get similar data than what it get already from Leela (best moves and sequences with %win rate). So I just started playing with it, in fact, I have the analysis/review part already quite working, I need more testing and integration to complete v0.7

I target to release v0.7 before the end of the month.

[johnsmith]

Thanks for all the great work, I'm sure many many more people appreciate it than just those who liked your post

What I meant in one of my previous posts was explained on the very, very simplified example (maybe even stupid example even though I'm dan player) in the attached images.

Image 1: first column is some kind of average of second and third column, while fourth is the probability of pro player spotting that move.
Image 2: if I understood GoReviewPartner correctly, then these columns are represented in this particular fashion as shown.

With current version I let GoReviewPartner analyzes the game for one hour, and then I have to switch to Leela UI to check number 3 (Net Win%). Which is small effort for me but still it's an effort What I'd like very much to have is number 3 in both current game window and variation window.

Cheers!

[pnprog]

Hi!

Image 1: first column is some kind of average of second and third column, while fourth is the probability of pro player spotting that move.
Image 2: if I understood GoReviewPartner correctly, then these columns are represented in this particular fashion as shown.

Below is the "dialogue" that will happen between GoreviewPartner and Leela "in the terminal" (this is the GTP protocol), for the same board position (I assumed a komi of 6.5) (blue are messages sent by GoReviewPartner, and red replies by Leela):

komi 6.5
=

play black t1
=

play black t19
=

play white d4
=

genmove white
MC winrate=0.609375, NN eval=0.732318, score=W+18.0
Nodes: 2869, Win: 79.88% (MC:64.70%/VN:85.92%), PV: D16 Q17 R4 P4 Q3
Nodes: 6744, Win: 80.93% (MC:64.06%/VN:87.64%), PV: D16 Q17 R4 P4 P3 O3 Q3 O4
Nodes: 9966, Win: 81.53% (MC:63.83%/VN:88.56%), PV: D16 Q17 Q4 R6 O3 P3
Nodes: 12727, Win: 81.49% (MC:63.84%/VN:88.50%), PV: D16 Q17 Q4 R6 O3 S4
Nodes: 15975, Win: 81.68% (MC:64.04%/VN:88.69%), PV: D16 F17 C14 Q17 R4 P4 Q5
Nodes: 18602, Win: 82.02% (MC:64.22%/VN:89.10%), PV: D16 Q17 Q4 R6 O3 R11 R10
Nodes: 23200, Win: 82.41% (MC:64.36%/VN:89.58%), PV: D16 Q17 Q4 R6 O3 R11 N16 O16
Nodes: 26451, Win: 82.47% (MC:64.31%/VN:89.70%), PV: D16 F17 C14 Q17 R4 P4 Q5 P5
Nodes: 30230, Win: 82.62% (MC:64.24%/VN:89.93%), PV: D16 F17 C14 Q17 R4 K16 Q14

D16 -> 29533 (W: 83.23%) (U: 64.20%) (V: 90.80%: 668) (N: 21.9%) PV: D16 F17 C14 Q17 R4 K16 Q14
D17 -> 1331 (W: 79.37%) (U: 64.12%) (V: 85.43%: 30) (N: 26.4%) PV: D17 Q16 R4 D15
Q16 -> 1170 (W: 81.01%) (U: 63.94%) (V: 87.80%: 28) (N: 13.3%) PV: Q16 Q4 D16 R14 O17 S16 R17 R11 S17
Q4 -> 672 (W: 79.06%) (U: 62.34%) (V: 85.71%: 14) (N: 14.2%) PV: Q4 D16 Q16 O3 O4
R4 -> 302 (W: 77.13%) (U: 65.55%) (V: 81.74%: 8) (N: 9.4%) PV: R4 D16 E16
C16 -> 250 (W: 77.23%) (U: 65.10%) (V: 82.06%: 8) (N: 7.6%) PV: C16 Q17 O17
====================================
29533 visits, score 83.23% (from 82.71%) PV: D16 F17 C14 Q17 R4 K16 Q14

33417 visits, 11858 nodes, 33417 playouts, 1336 p/s

= D16

The interesting part is:

D16 -> 29533 (W: 83.23%) (U: 64.20%) (V: 90.80%: 668) (N: 21.9%) PV: D16 F17 C14 Q17 R4 K16 Q14

So the information currently displayed are:

• Black/white win probabilities: W: 83.23%
• Number of play-outs: 29533
• Neural network value: N: 21.9%

So we are left with those 2 informations: (U: 64.20%) and (V: 90.80%: 668) and I don't know what they mean. Is (V: 90.80%: 668) the one you want? (to be understood as 100%-90.80% for white maybe? so 9.20%)

[johnsmith]

Yes, (V: 90.80%: 668) is what I'd like it is the same like those parameters above:
W => Win
U => Monte Carlo (MC)
V => Value Network (VN)

The thing is, similar to this example, that MC can sometimes show evaluation of 50% while value net can be as high as 75%. So MC might trick you the game is even, while value net tells you how things really are. The same thing goes for Win since it takes MC into account...
So you're right - since it is white's turn it means white is leading by 90.8%, and black is losing by 100-90.8=9.2%. That number 668 I have no idea what it means

edit:

to be understood as 100%-90.80% for white maybe? so 9.20%

for black I think...

[roy7]

V is the value network win rate estimate. It moves much quicker, and generally is more accurate, than the U (Monte Carlo) win rate. However it is subject to blind spots where it makes errors (like the policy network does) so the author uses a weighted average W of the U and V values (bit heavier weight on the Monte Carlo value) for the decision making process when Leela does tree search.

N is the policy network based odds a pro player would choose that move. The tree search is seeded and weighted based on the N values. If a move has an N value of 0, Leela won't consider it, which can lead to the rare "blunder" moves where Leela just can't see a move.

The number after the V % is (I think) the number of nodes evaluated by the value network in that variation. The NN evaluation, even with GPU, is much slower than the tree search itself can explore.

[pnprog]

Ok, so it seems I was quite wrong regarding all those values
I will fix that quickly and add all those values in a v0.6.1 that I will try to release tomorrow.

[pnprog]

Hi!

When trying on a 9x9 games, I got this type of data from Leela instead:

C3 -> 17924 (U: 89.10%) (R: 87.78%: 33386) (N: 144.9%) PV: C3 C4 C5 B3 B4 B5 D4 B6

So:

• 17924: Number or play-outs
• (U: 89.10%): I guess this is still the MonteCarlo search result
• (R: 87.78%: 33386): Do you know what that one would be?
• (N: 144.9%): Apparently, this is: Eval: displayed in non-19x19 games instead of Net Prob%. A dimensionless number expressing the likelihood of the move being good, based on a pattern database and the moves' features.

Best regards,

[johnsmith]

9x9 is really puzzling.