The tool is awesome, but has some unused potential, to make it even better...A switch from Engine A to B by movenumber is not as good as by number of pieces on the board (in pawns, added to one number). But movenumber is not so bad as it seems, because of Lc0's bad time-management: When playing with increment, lc0 uses the whole basetime until move 60 most of the games. So, a switch to a AB-engine some moves before 60 makes always sense.
As I reported here before, Chimera works fine in FritzGUI 17 (pondering works, too (!)). Now, I can report, it works fine in cutechess-cli, too. I started a testrun for my NBSC-Advanced Armageddon testing vs. Stockfish:
https://www.sp-cc.de/nn-vs-sf-testing.htm
Chimera (first 52 moves Lc0 0.24.1 LS 14.3, then Komodo 14 for the rest of the games (Mention, that 4 moves of each NBSC-opening lines are just for making the short castle of black illegal, so 48 moves are played, only, when Chimera switches to Komodo 14). And Chimera switches 1 or 2 moves (depends on played color) too late to the Engine B, so Lc0 plays the first 50 moves (plus/minus one move)).
My config-file of Chimera (chimera.xml) looks like this:
(Important: You have to change SwitchingOverheadSeconds="1" manually with an editor (default is 30, which is way too much for short timecontrols!!)).
With that configuration, the first 48 of 500 games in my testrun were played without any timelosses, stalls or other problems with cutechess-cli.
Code: Select all
<ChimeraUciConfig SwitchingOverheadSeconds="1" ShowThreads="false" ShowHash="false">
<Name>chimera</Name>
<Engine FromMove="0">
<Exe>C:\Cutechess\Engines\Chimera\lc0.exe</Exe>
<Arguments />
<Username />
<Password />
<Param Name="Backend">cudnn-fp16</Param>
<Param Name="NNCacheSize">5000000</Param>
<Param Name="OutOfOrderEval">False</Param>
<Param Name="MoveOverheadMs">1000</Param>
</Engine>
<Engine FromMove="52">
<Exe>C:\Cutechess\Engines\Chimera\komodo-14-64bit-bmi2.exe</Exe>
<Arguments />
<Username />
<Password />
<Param Name="Threads">11</Param>
<Param Name="Ponder">False</Param>
<Param Name="Hash">2048</Param>
<Param Name="Overhead ms">1000</Param>
<Param Name="Contempt">0</Param>
</Engine>
</ChimeraUciConfig>
And mention, that Komodo 14, which plays from move 50 vs Stockfish is
a) around 100 Elo weaker than Stockfish in ratinglists
b) Stockfish has +50% more thinkingtime in my testruns, in order to get a Leela-Ratio of 1.0: Lc0/Chimera: 2'+1'', SF has 3'+1.5'' thinking-time. That is a clear disadvantage for Chimera, because Komodo uses the CPU (of course), like SF and because of this, SF has +50% more time than Komodo 14, when Komodo 14 takes over at move 50...
Final result in 2-3 days. Stay tuned!
(When the testrun is finished, I will try the same run again, with Chimera switching to Komodo 14 10 moves earlier...Closer to the movenumber, K.Laskos suggests (35). Lets see, what is better)
With Chimera, other engine-setups are possible (and interesting), of course:
What about switching from one lc0 to another, when the second lc0 uses an endgame NN ?
https://github.com/dkappe/leela-chess-w ... ndgame-Net
Could be interesting, too....
