Decades ago, I enjoyed training against dedicated chess units at "training levels" (still do). I mean me playing the machines to prepare for a tournament. Training levels set the engine to a fixed ply search depth. Many of them like my trusty CXG Super Crown have the lowest training
level at a fixed depth of 1 ply search and each greater training level added 1 ply to the search. If you make a mistake, you know just how deep you didn't search!
Many of the machines could go to 8 ply some more. Some research was published in 1990 suggesting that each ply was worth about 200 Elo. Today some claim 75 Elo per ply, but I suspected that was way off for lower plies and that the rating to ply curve is nonlinear. So, I tested Ares at several different
fixed plies against other engines at timed TC in blitz. Using the CCRL blitz ratings I've an approximation of Ares playing strength per ply.
The results indicate a 1 ply search (with Qsearch enabled) is around 900 Elo. Of this would change based on the strength of an engine's position eval. The Elo increase per ply is about 150 until it reached 7 ply. The table shows CCRL blitz ratings of the other engines and g-ratings (guess) of Ares.
I plan to put more engines in the list so that there is an unencumbered engine every 100 points instead of 200 points. At that point, I'll do some more
serious math for rating estimates.
It could be interesting to run the same tests again with Qsearch disabled. In certain positions, Qsearch really helps and others not.
The table can be found at http://aigames.net/Research/fdseries.html
Ares playing strength at various fixed depth levels
Moderators: hgm, Rebel, chrisw
-
- Posts: 2056
- Joined: Mon Mar 13, 2006 2:31 am
- Location: North Carolina, USA
-
- Posts: 5588
- Joined: Wed Sep 05, 2018 2:16 am
- Location: Moving
- Full name: Jorge Picado
Re: Are playing strength at various fixed depth levels
Try training with Honey by setting it at your Elo strength, if you use Arena GUI click Engine and click engine1 and click configure and select your level https://github.com/MichaelB7/Stockfish/releases/tag/14CRoberson wrote: ↑Thu Jun 17, 2021 12:10 am Decades ago, I enjoyed training against dedicated chess units at "training levels" (still do). I mean me playing the machines to prepare for a tournament. Training levels set the engine to a fixed ply search depth. Many of them like my trusty CXG Super Crown have the lowest training
level at a fixed depth of 1 ply search and each greater training level added 1 ply to the search. If you make a mistake, you know just how deep you didn't search!
Many of the machines could go to 8 ply some more. Some research was published in 1990 suggesting that each ply was worth about 200 Elo. Today some claim 75 Elo per ply, but I suspected that was way off for lower plies and that the rating to ply curve is nonlinear. So, I tested Ares at several different
fixed plies against other engines at timed TC in blitz. Using the CCRL blitz ratings I've an approximation of Ares playing strength per ply.
The results indicate a 1 ply search (with Qsearch enabled) is around 900 Elo. Of this would change based on the strength of an engine's position eval. The Elo increase per ply is about 150 until it reached 7 ply. The table shows CCRL blitz ratings of the other engines and g-ratings (guess) of Ares.
I plan to put more engines in the list so that there is an unencumbered engine every 100 points instead of 200 points. At that point, I'll do some more
serious math for rating estimates.
It could be interesting to run the same tests again with Qsearch disabled. In certain positions, Qsearch really helps and others not.
The table can be found at http://aigames.net/Research/fdseries.html
Who is 17 years old GM Gukesh 2nd at the Candidate in Toronto?
https://indianexpress.com/article/sport ... t-9281394/
https://indianexpress.com/article/sport ... t-9281394/