Pre-release study about draw rates and diminishing returns

[Rebel]

Pre-release study about draw rates and diminishing returns.



I don't think these numbers are a big surprise for insiders, but it's good to have it researched.

I need about 2-3 more days to finish the last (time consuming) matches but I think the picture is clear enough.

http://rebel13.nl/rebel13/time-odds-matches.html

[Jouni]

Yes with NNUE gains from time, threads or tablebases are only about 50% of old rule of thumb. But is the reason NNUE or the improved strength ?

[Rebel]

Yes with NNUE gains from time, threads or tablebases are only about 50% of old rule of thumb. But is the reason NNUE or the improved strength ?

Can you rephrase the question ?

[Rebel]

Still low draw rates with 20 cores.

Maybe unbalanced -- but playable -- positions is the future, at least for the entertaining part.

[amanjpro]

Yes with NNUE gains from time, threads or tablebases are only about 50% of old rule of thumb. But is the reason NNUE or the improved strength ?

Can you rephrase the question ?

I believe the question is:

We observe that: the allotted time, number of threads and using EGTB are less effective nowadays, maybe by the factor of 50% of what we once thought it is... What could be the reason: is it NNUE? or the engines are too powerful nowadays that they don't need much help from those factors

[Rebel]

Tester Mark Young did an interesting time-odds experiment.

Stockfish 14 (32 threads, Hash 4 GB)
Stockfish 14 (32 threads, Hash 32 GB)
Time Control - 5m+5s vs 60m+60s

Result after 12 games

Code: Select all

Result:
----------------------------------------------------------------------------------
  #  name                  games    wins   draws  losses   score    elo    +    -
  1. Stockfish 14 5m+5s       12       0      12       0       6      0   63   63
  2. Stockfish 14 60m+60s     12       0      12       0       6      0   63   63

  #  name                    nodes/m         NPS  depth/m   time/m    moves     time
  1. Stockfish 14 5m+5s      263911K    29566552     52.6      8.9     65.0    580.2
  2. Stockfish 14 60m+60s   3341118K    30431351     67.2    109.8     64.9   7127.3   

Meanwhile there are now 16 draws.

To follow the experiment go to - https://prodeo.actieforum.com/t591-stoc ... sh-14-5m5s

[Rebel]

Meanwhile I copied this (great) idea for Dragon 2.5 but.... with the gambit positions.

One game takes about 1½ - 2 hours.

http://rebel13.nl/b/grl.htm

[lkaufman]

Meanwhile I copied this (great) idea for Dragon 2.5 but.... with the gambit positions.

One game takes about 1½ - 2 hours.

http://rebel13.nl/b/grl.htm

I think that results will depend heavily on the positions chosen. If they evaluate (on a reasonable search) near zero, you'll get almost all draws. If the eval is half a pawn or more favorable to one side, then the version using the long time limit will win many of the games where it starts with the favorable side of this opening, and draw the others.

[Rebel]

Meanwhile I copied this (great) idea for Dragon 2.5 but.... with the gambit positions.

One game takes about 1½ - 2 hours.

http://rebel13.nl/b/grl.htm

I think that results will depend heavily on the positions chosen. If they evaluate (on a reasonable search) near zero, you'll get almost all draws. If the eval is half a pawn or more favorable to one side, then the version using the long time limit will win many of the games where it starts with the favorable side of this opening, and draw the others.

I put my money on another SF issue, like with the knight odds.

Mark has started the same match now with Dragon 2.5



And so I can stop after 8 games

Code: Select all

No. Name               Win Draw Loss Unf.  Score Games       %
--------------------------------------------------------------
  1 Dragon2.5 60m+60s   +3   =5   -0   *0    5.5     8   68.8%
  2 Dragon2.5 5m+5s     +0   =5   -3   *0    2.5     8   31.2%

3 wins, 5 draws. SF14 got only 1 win out of 23 games.