Uncomment resign code to allow lczero to resign #418

jjoshua2 · 2018-04-23T00:18:20Z

removed comments

whitespace is better now hopefully?

I don't know what rootstate was, probably should have been root

killerducky · 2018-04-23T00:29:43Z

I don't think UCI engines resign, it's up to the GUI to do that.

jjoshua2 · 2018-04-23T00:30:49Z

This is like how leela go does it, it's for self play only...

position has game ply...

killerducky · 2018-04-23T00:39:01Z

LZGo uses GTP, which specifies how to resign. UCI does not.

How is the client going to know who won? You will need a custom UCI command. I think it's better to have the client parse the UCI output and look at the cp score. This way we don't need custom UCI commands and lczero code doesn't have to change.

killerducky · 2018-04-23T00:41:26Z

Oh now I remember client doesn't speak UCI during the game, it just says "train" and the entire game runs. So I guess it has to be done similar to how you are doing it.

killerducky · 2018-04-23T00:47:09Z

I think you did some resign analysis? Can you post it here?

keeps it from doing a1+ as a move_none

5% winrate is a better initial resign target

Tilps · 2018-04-23T02:24:10Z

src/UCTSearch.cpp

@@ -196,18 +196,16 @@ Move UCTSearch::get_best_move() {
        return bestmove;
    }

-    // should we consider resigning?
-    /*
+    // should we consider resigning?    
       float bestscore = m_root->get_first_child()->get_eval(color);
       int visits = m_root->get_visits();
    // bad score and visited enough
    if (bestscore < ((float)cfg_resignpct / 100.0f)
        && visits > 500


This visits > 500, seems rather specific to the fact that training is run at 800 visits.
Also this whole logic should be protected in some way to ensure its only run during training?

That 500 visits was there. I didn't put that in. It makes sense to have a threshold to keep it from resigning when it can't get enough playouts to determine if its a good move though. Default resignpct should probably be 0, then it won't resign, unless overridden...

I definitely think there needs to be a minimum visit threshold, I just wonder if it should be calculated rather than a constant. Its much easier to get 500 visits to a 'best' move which is actually almost tied three ways, if the actual number of visits is 1600, rather than 800. This is relevant if the other moves were discovered late and have good win rates but haven't quite caught up to the leader by the 1600 visits. Temp is actually more likely to choose a move other than this one, since its less than half the visits.
Maybe > half the visits on the one option, and also > 500 to have confidence that the eval is reasonably calculated. (Or maybe all options with > 500 visits must agree to resign? Or the weighted majority of options with > 500 visits? I don't know what is best.)

This condition was removed upstream. It dates back to the time when the score was estimated from Monte Carlo playouts, and doesn't make a lot of sense with a strong neural network evaluator.

jjoshua2 · 2018-04-23T02:38:14Z

Note: playMatch does not understand the NONE_MOVE so it will have a Error decoding:
easy way to fix that is just not use resigns in matches like we currently do. they dont have the temperature problem anyway as much

jjoshua2 · 2018-04-23T02:43:29Z

With T=1 this is a major time win.
Completed 14 games in 3h3m cpu time vs 17 games in 1h51m with 5% resign.
17.4 min/game vs 6.5 min/game. So we could maybe double our playouts here...

killerducky · 2018-04-23T02:52:15Z

Please see the current LZGo codebase, the 500 magic number is gone, along with many other changes.

killerducky · 2018-04-23T03:00:56Z

I saw in the chat shyeel talking about doing some code for statistics. First, see for reference https://github.com/gcp/leela-zero/blob/master/scripts/resign_analysis/resign_analysis.py

The main thing to measure is "incorrect resigns" and "moves saved by resigning". You need to analyze self-play games that have resign disabled. Calculate who would have resigned, and count how often it was the wrong side (incorrect resign). Calculate how many moves were saved. This is our cost/benefit analysis.

Tilps · 2018-04-23T03:56:12Z

The LZGo codebase appears not to use temperature after x moves - so they can do proper resign analysis easily. Also resignation is therefore mostly about improving games per hour, and possibly focusing the engine to learn less about deep lost endgames that no one is actually going to play out in practice. Its not about providing a temperature reducing effect.

jkiliani · 2018-04-24T20:24:47Z

Yes, Leela Zero uses the temperature parameters directly from the Alphago Zero paper, which uses t=1 for the first 30 moves and t->0 for the rest of the game. In fact, fractional temperature is not implemented at all in Leela Zero. They don't really need it since the move space is so much bigger, and symmetry application provides another source of randomness. I don't really like the sharp cutoff where moves up to no. 30 can include any blunder, while those from move 31 don't, but arguably it works.

killerducky · 2018-04-26T03:18:50Z

float bestscore = m_root->get_first_child()->get_eval(color);
I think we should use m_root get_eval, because that's what cp eval was tuned for, and that's what the resign analysis I did was based on. It's an arbitrary choice because even plies will be biased one way and odd plies will be biased the other way. Best just stick to the one we've got.

cfg_resignpct should default to something that means never resign. I guess a magic -1, which actually works with the current code as is. client can override this.

I think we can just remove the visits qualifier.

won't resign by default now

no visit limit now since we do 800 playouts anyway and this is not for uci

killerducky · 2018-05-01T23:03:19Z

src/UCI.cpp

+      bh.do_move(move);
+    } else {
+       return bh.cur().side_to_move() == WHITE ? -1 : 1;
+    }


@glinscott does this part look ok to you? Making MOVE_NONE mean resign?

killerducky · 2018-05-01T23:04:42Z

@glinscott I put one question in the code diff, can you take a look? If that is ok I think we should pull this and then the clients will be ready for when the server starts sending the --resign N switch. By default they will act the same as before so this should be safe.

jjoshua2 added 3 commits April 22, 2018 20:17

Uncomment resign code to allow lczero to resign

0e42e78

Update UCTSearch.cpp

d4bef70

whitespace is better now hopefully?

Update UCTSearch.cpp

8a2cec8

I don't know what rootstate was, probably should have been root

Update UCTSearch.cpp

03803c7

position has game ply...

jjoshua2 added 2 commits April 22, 2018 21:10

Update UCI.cpp

19ba541

keeps it from doing a1+ as a move_none

Update Parameters.cpp

b759453

5% winrate is a better initial resign target

killerducky changed the base branch from master to next April 23, 2018 01:38

Tilps reviewed Apr 23, 2018

View reviewed changes

This was referenced Apr 24, 2018

Update resign_analysis script for chess. #427

Merged

A faster winning condition #429

Open

Update Parameters.cpp

34a8f5f

won't resign by default now

killerducky changed the base branch from next to master May 1, 2018 15:31

killerducky changed the base branch from master to next May 1, 2018 15:31

jjoshua2 and others added 3 commits May 1, 2018 11:38

Update UCTSearch.cpp

3a8b087

no visit limit now since we do 800 playouts anyway and this is not for uci

Merge remote-tracking branch 'upstream/next' into HEAD

0cc8dbd

Minor cleanups

b3ea1de

killerducky reviewed May 1, 2018

View reviewed changes

killerducky merged commit 42c623c into glinscott:next May 2, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uncomment resign code to allow lczero to resign #418

Uncomment resign code to allow lczero to resign #418

jjoshua2 commented Apr 23, 2018

killerducky commented Apr 23, 2018

jjoshua2 commented Apr 23, 2018

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

Tilps Apr 23, 2018

jjoshua2 Apr 23, 2018

Tilps Apr 23, 2018

gcp Apr 24, 2018

jjoshua2 commented Apr 23, 2018

jjoshua2 commented Apr 23, 2018 •

edited

Loading

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

Tilps commented Apr 23, 2018

jkiliani commented Apr 24, 2018

killerducky commented Apr 26, 2018

killerducky May 1, 2018

killerducky commented May 1, 2018

Uncomment resign code to allow lczero to resign #418

Uncomment resign code to allow lczero to resign #418

Conversation

jjoshua2 commented Apr 23, 2018

killerducky commented Apr 23, 2018

jjoshua2 commented Apr 23, 2018

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

Tilps Apr 23, 2018

Choose a reason for hiding this comment

jjoshua2 Apr 23, 2018

Choose a reason for hiding this comment

Tilps Apr 23, 2018

Choose a reason for hiding this comment

gcp Apr 24, 2018

Choose a reason for hiding this comment

jjoshua2 commented Apr 23, 2018

jjoshua2 commented Apr 23, 2018 • edited Loading

killerducky commented Apr 23, 2018

killerducky commented Apr 23, 2018

Tilps commented Apr 23, 2018

jkiliani commented Apr 24, 2018

killerducky commented Apr 26, 2018

killerducky May 1, 2018

Choose a reason for hiding this comment

killerducky commented May 1, 2018

jjoshua2 commented Apr 23, 2018 •

edited

Loading