Play move with the highest lower confidence bound

[Ttl]

Same as leela-zero/leela-zero#2290. Some features still missing from the LZ PR: minimum visit ratio and avoiding pruning of the best LCB move in time management.

800 node test with network 41724 (MinibatchSize=32 and 5 piece TB)
Score of lc0_lcb vs lc0_master: 85 - 62 - 253  [0.529] 400
Elo difference: 20.00 +/- 20.61, LOS: 97.11 %, DrawRatio: 63.2 %

[killerducky]

I think you'll need something similar to this in meson.build:

deps += [ dependency('boost', modules : ['thread', 'system' ], static : true) ]

[Ttl]

Oh, boost isn't used anywhere. Somehow it works on my machine. Probably no point in adding it just for this one function call.

[Ttl]

I clopped ci_alpha and min_ratio parameters at 800 nodes with T50 network against master branch:

X-axis: GammaParameter cialpha 1e-7 1e-2
Y-axis: LinearParameter minimum-lcb-n-ratio 0.05 0.95

[ddobbelaere]

Nice idea, excited to see it works so well!

[ddobbelaere]

Technically this is the standard error (standard deviation divided by sqrt(N)). Maybe stderr is a better variable name.

Or is this the standard deviation of something in another context?

[Ttl]

I was thinking of the estimated distribution for the mean which has this standard deviation. Standard error seems to be correct too.

[ddobbelaere]

Okay, that seems logical, I didn't think about it that way :).

[Ttl]

Some testing at high nodes, nothing conclusive. TC 30+0.5s. Ryzen 1700X and RTX2080. SF with 16 threads.

   # PLAYER              :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)
   1 lc0_41724_master    :    24.7   15.8   274.0     512    54      67
   2 lc0_41724_lcb       :    17.8   15.4   269.0     512    53      96
   3 stockfish_dev       :     0.0    9.8   481.0    1024    47     ---

[borg323]

Quick patch to fix appveyor builds: borg323@bad4503

[zz4032]

TC 4.8s/game+0.02s/move (~900 nodes/move)

   # PLAYER       :  RATING  ERROR  LOS(%)   GAMES  DRAWS(%)
   1 lc0_PR817    :    47.4   12.5   100.0    1000      60.7
   2 lc0          :     0.0   ----     ---    1000      60.7

White ELO advantage = 56.5
Draw rate (equal opponents) = 64.3 %

TC 17.04s/game+0.071s/move (~6000 nodes/move)

   # PLAYER       :  RATING  ERROR  LOS(%)   GAMES  DRAWS(%)
   1 lc0          :     0.0   ----    60.0    1000      69.0
   2 lc0_PR817    :    -1.5   11.2     ---    1000      69.0

White ELO advantage = 68.1
Draw rate (equal opponents) = 73.9 %

[zz4032]

Tried retuning this PR and noticed that at lower nodes/move I'm getting a weighted plot similar to yours:

But with increased time control (~12000 nodes/move) the optimum area shifts and split's in two:

Clearly a reason why it doesn't perform as expected at LTC.

[Naphthalin]

As far as I know, this is still relevant, and not solved by time manager or so. Might be a candidate for @kiudee tuning it once updated, and merged if useful.

[Naphthalin]

Closing this for a combination of several reasons:

1. implementation requires boost
2. too many search conflicts
3. Analyse Mode + minimax/mcts hybrid #1543 contains a more recent LCB implementation
4. implementation of Extracting parts of Lc0's search into classes would help future development #1734 will allow a much easier reimplementation, compared to updating this branch.