tests of stockfish1.9.1 against new stockfish with a single core

[jkominek]

For your depth vs depth experiments my first step is to calibrate SF 17.1 run at fixed depths against fixed nodes / move.

Code: Select all

Stockfish 17.1 Hash 512 Threads 1 Depth 12 vs. Stockfish 17.1 Hash 512 Threads 1 Nodes 32768 [+100, =384, -164] -34.4 Elo
Stockfish 17.1 Hash 512 Threads 1 Depth 11 vs. Stockfish 17.1 Hash 512 Threads 1 Nodes 16384 [+134, =365, -149]  -8.0 Elo
Stockfish 17.1 Hash 512 Threads 1 Depth 10 vs. Stockfish 17.1 Hash 512 Threads 1 Nodes 8192  [+157, =335, -156]  +0.5 Elo

For SF 1.91 it would take too long for direct measurements. Instead I decided to use the median number of nodes per move. I went with median instead of mean because when running with fixed depth Stockfish can get stuck at times, shifting the average upward due to some very large node counts. It's not just a problem of the past, either. Annoyingly I find the most recent stockfish-dev can get stuck on fixed depth searches.

[url][/ur]

Drawing the comparisons:

• Stockfish 17.1 d=10 lines up with Stockfish 1.91 4 Mnodes/mv ~ d=21
  Stockfish 17.1 d=12 lines up with Stockfish 1.91 256 Mnodes/mv ~ d=29, one notch above your ending point of d=28.