LZ that does not escape difficult ladder is slightly(54%) stronger

[yssaya]

If ladder search find its dead over 7 plies, escaping move probability is changed to 1/1000000.

"not escape difficult ladder" winrate is 0.541 (433 wins / 800 games) against normal LZ 0.17.

Source is here. (forked from latest LZ)
https://github.com/yssaya/leela-zero-ladder
I use Henrik Forsten(Ttl) ladder code.
https://github.com/Ttl/leela-zero/tree/liberty_planes_se2

For example, in left figure, ladder search needs 42 plies to understand Black E7 is killed.
So Black E7 probablity is changed from 0.55895 to 0.00000055895.
In right figure, ladder search needs 6 plies. So program does nothing and LZ plays escape move Black N17.

   a b c d e f g h j k l m n o p q r s t        a b c d e f g h j k l m n o p q r s t
19 . . . . . . . . . . . . . . . . . . . 19  19 . . . . . . . . . . . . . . . . . . . 19
18 . . . . . . . . . . . . . O . . . . . 18  18 . . . . . . . . . . . . . O . . . . . 18
17 . . . . . . . . . . . . . X O O O . . 17  17 . . . . . . . . . . . . . X O O O . . 17
16 . . . X . . . . . + . . . . X X X . . 16  16 . . . X . . . . . + . O X(O)X X X . . 16
15 . . . . . . . . . . . . . . . . . . . 15  15 . . . . . . . . . . O X X O . . . . . 15
14 . . . . . . . . . . . . . . . . . . . 14  14 . . . . . . . . . O X X O . . . . . . 14
13 . . . . . . . . . . . . . . . . . . . 13  13 . . . . . . . . O X X O . . . . . . . 13
12 . . . . . . . . . . . . . . . . . . . 12  12 . . . . . . . O X X O . . . . . . . . 12
11 . . . . . . . . . . . . . . . . . . . 11  11 . . . . . . O X X O . . . . . . . . . 11
10 . . . + . . . . . + . . . . . + . . . 10  10 . . . + . O X X O + . . . . . + . . . 10
 9 . . . . . . . . . . . . . . . . . . .  9   9 . . . . O X X O . . . . . . . . . . .  9
 8 . . .(O). . . . . . . . . . . . . . .  8   8 . . . O X X O . . . . . . . . . . . .  8
 7 . . O X . . . . . . . . . . . . . . .  7   7 . . O X X O . . . . . . . . . . . . .  7
 6 . . X O O . . . . . . . . . . . . . .  6   6 . . X O O . . . . . . . . . . . . . .  6
 5 . . X . . . . . . . . . . . . . . . .  5   5 . . X . . . . . . . . . . . . . . . .  5
 4 . . X O . . . . . + . . . . . O . . .  4   4 . . X O . . . . . + . . . . . O . . .  4
 3 . . . . . X . . . . . . . . . . . . .  3   3 . . . . . X . . . . . . . . . . . . .  3
 2 . . . . . . . . . . . . . . . . . . .  2   2 . . . . . . . . . . . . . . . . . . .  2
 1 . . . . . . . . . . . . . . . . . . .  1   1 . . . . . . . . . . . . . . . . . . .  1
   a b c d e f g h j k l m n o p q r s t        a b c d e f g h j k l m n o p q r s t

http://www.yss-aya.com/cgos/viewer.cgi?19x19/SGF/2019/09/11/599617.sgf

Ignoring short depth is effective when killing eye.

4 . . O + X X X .
3 . X O O O X . .
2 . O X . . . . .    White A2 is killed by 2 plies search, but needed to kill eye.
1 .(X). . . . . .
  a b c d e f g h

I only tested 400 visit/move. It might be even or weak on longer game or time patiry.

Test match is 800 games. First 10 moves are from opening book, pro(gogod) games.
Symmetry position are deleted. Black and white is alternate per a opening. Weight is LZ 244.

Test command is like this.
$ java -jar ~/go/gogui/bin/gogui-twogtp.jar -black "~/go/ladder/build/leelaz17 -v 400 -r5 -g -q -d --noponder -w ~/go/leela-zero-0.15/autogtp/networks/2baef15960bf40ebf613f5b0b738d41b89646332403d105e11f8623d54cce792" -white "~/go/ladder/build/leelaz -v 400 -r5 -g -q -d --noponder -w ~/go/leela-zero-0.15/autogtp/networks/2baef15960bf40ebf613f5b0b738d41b89646332403d105e11f8623d54cce792" -games 2000 -size 19 -alternate -sgffile memo -auto -komi 7.5 -openings "~/go/ladder/opening/19x19_gogod_2014_1000"

[l1t1]

nice found

[yssaya]

As a result of increasing the number of games a little more, winrate was 0.529. (800 visit/move, 2000 games)
Here is leelaz.exe for Windows.
https://github.com/yssaya/leela-zero-ladder/releases/tag/v0.17_ladder
This will work if you replace leelaz.exe with a normal one.

I also tested "allow escaping when ko" and "Do not chase 'not ladder'".
But both are not significantly stronger.

Result against normal LZ. weight is 244. 800 visit/move. using opening book for first 10 moves.

"Do not escape difficult ladder"  1058 wins  942 losses (2000games) winrate 0.529
"allow escaping when ko"           998 wins 1002 losses (2000games) winrate 0.499
"Do not chase 'not ladder'"       1024 wins  976 losses (2000games) winrate 0.512

A sample of "Do not chase 'not ladder'".
http://www.yss-aya.com/cgos/viewer.cgi?19x19/SGF/2019/10/08/609399.sgf&110

Test command is like this.
$ java -jar ~/gogui/gogui-twogtp.jar -black "~/go/ladder/leelaz17 -v800 -r5 -g -q -d --noponder -w ~/go/networks/244_2bae" -white "~/go/ladder/leelaz -v800 -r5 -g -q -d --noponder -w ~/go/networks/244_2bae" -games 2000 -size 19 -alternate -sgffile memo -auto -komi 7.5 -openings "~/go/ladder/opening/19x19_gogod_2014_1000"

Average searching speed of 10 positions is as follows. (a 2080 Ti, 10000 playouts)

932 n/s  ... Normal
944 n/s  ... "Do not escape difficult ladder"

"Do not escape difficult ladder" is slightly faster. But it should be measurement error.
There may be almost no decrease in speed.

[l1t1]

testing validation -k 768 -n \768.txt -o "-g -v 100 -r 5 -w" -n \768.txt -o "-g -v 100 -r 5 -w" -- leelaz_lad.exe leelaz.exe

94 wins, 82 losses
176 games played.
Status: 0 LLR 0.314042 Lower Bound -2.94444 Upper Bound 2.94444

[yssaya]

I have tested against KataGo.

Result against KataGo 1.2 g104-b20c256. LZ weight is 247. Both players use 800 visit/move. First 10 moves are from opening book.

"Normal Leela Zero"               978 wins 1022 losses (2000games) winrate 0.489, ELO=  -7.6
"Do not chase 'not ladder'"      1042 wins  958 losses (2000games) winrate 0.521, ELO= +14.6

"Do not chase 'not ladder'" is +22.2 ELO stronger than normal LZ.
LZ often lost to KataGo by "chasing not ladder", so I thought this option would work.
But it wasn't.

[nemja]

I don't think adding ladder code without retraining the net is anything close to what other bots have.

[wonderingabout]

i suggest you try with --batchsize 1 -t 2
as Naphtalin showed on discord, these settings have much stronger ladder performance

[portkata]

Can you add alreadydone's "dynamic komi" code to your engine so we can play with 0 komi? Thanks!