UCI_Elo 1500 engines and FIDE Elo

lkaufman · Post by **lkaufman** » Sat Apr 24, 2021 10:41 pm

pedrox wrote: ↑Sat Apr 24, 2021 3:56 pm
lkaufman wrote: ↑Sat Apr 24, 2021 1:34 am
pedrox wrote: ↑Sat Apr 24, 2021 12:05 am
MikeB wrote: ↑Thu Apr 22, 2021 5:09 am
lkaufman wrote: ↑Wed Apr 21, 2021 6:24 am
MikeB wrote: ↑Tue Apr 20, 2021 5:25 pm
Ferdy wrote: ↑Tue Apr 20, 2021 12:02 pm Maia Lichess 1546 has a blitz Lichess rating of around 1546. This engine is run with maia-1100.pb.gz net at node 1 on a gauntlet test below.

The formula to get FIDE Elo from Lichess Blitz Elo is,
Code: Select all
FIDE rating = 65 + [0.87 x Lichess Blitz Rating]
based from the regression in 2019.

Maia FIDE Elo = 65 + 0.87 x 1546 = 1410

Maia gauntlet bullet at TC 60s+100ms and is set at FIDE Elo 1410.
Code: Select all
   # Player                                    :  FIDE Elo  Points  Played  Points(%)
   1 Rybka v2.3.2a UCI_Elo 1500                :      2217    99.0     100       99.0
   2 Minic 2.51 UCI_Elo 1500                   :      2145    98.5     100       98.5
   3 Cheese 2.1 UCI_Elo 1500                   :      2094    98.0     100       98.0
   4 Honey X5i UCI_Elo 1500                    :      2054    97.5     100       97.5
   5 Cheng 4.39 UCI_Elo 1500                   :      2054    97.5     100       97.5
   6 Ufim v8.02 UCI_Elo 1500                   :      2021    97.0     100       97.0
   7 Amyan 1.72 UCI_Elo 1500                   :      1910    94.5     100       94.5
   8 Wasp 4.5 UCI_Elo 1500                     :      1840    92.0     100       92.0
   9 MadChess 3.0 UCI_Elo 1500                 :      1817    91.0     100       91.0
  10 Deuterium v2021.1.38.29.4 UCI_Elo 1500    :      1753    87.5     100       87.5
  11 Rhetoric 1.4.3 UCI_Elo 1500               :      1737    86.5     100       86.5
  12 Discocheck 5.2 UCI_Elo 1500               :      1730    86.0     100       86.0
  13 Arasan 21.1 UCI_Elo 1500                  :      1716    85.0     100       85.0
  14 Rodent IV 022 UCI_Elo 1500                :      1716    85.0     100       85.0
  15 Deuterium v2019.2.37.73 UCI_Elo 1500      :      1599    74.5     100       74.5
  16 Danasah 7.9 UCI_Elo 1500                  :      1599    74.5     100       74.5
  17 Stockfish 13 UCI_Elo 1500                 :      1539    67.5     100       67.5
  18 Hiarcs 14 UCI_Elo 1500                    :      1449    55.5     100       55.5
  19 Maia Lichess 1546                         :      1410   312.5    1900       16.4
  20 Deuterium v2014.2.33.276 UCI_Elo 1500     :      1171    20.5     100       20.5
The closest engines at FIDE Elo 1500 within 100 Elo are Stockfish, Hiarcs, Maia, Danasah and Deuterium v2019.2.37.73.
Looks like several engines are tuning to CCRL 1500 which is significantly stronger than 1500 FIDE . about 1890 +/- 100
Does this last sentence mean that on average, an engine with a CCRL blitz rating of 1500 should be an even match for a human with a FIDE rating of about 1890 if they are playing blitz chess? If so what is the time limit assumed for the human vs engine match? I suppose it is based on the average blitz game played between humans and engines on Lichess, is that correct? If so can you estimate the average time limit of such human vs engine games? Time limit makes a huge difference in engine vs human play; a human might be hundreds of elo stronger against engines at game in ten minutes than at game in three minutes. This estimate of about four hundred elo gap between CCRL and FIDE ratings in blitz might be about right at some time control in that range.
I discovered this prior to LiChess became big, but yes I am talking Blitz Chess - around 3 3 maybe. I' am a 1600 player, and few years, when playing Shadow Blue ( about 1700 CCRL) for calibrating my Honey engines, - I realized immediately that it was not 1700 USCF, SO I dug deeper and came across this post with a lot of good info

http://talkchess.com/forum3/viewtopic.p ... 2&#p661409

So in my Honey Engines I let the user to choose CCRL (for the programmers) or FIDE ( for humans).

Convert CCRL to FIDE:

(CCRL-Ra * 10) / 7) - 1200 = FIDE-RA

Convert FIDE to CCRL:

(FIDE-Ra+1200)*7/10=CCRL-Ra

.. at 2800 , they are about the same ... (2800 *10 /7 ) - 1200 = 2800

Most my Elo testing of engines have been around 3 3. I don't play well faster than that ,and longer times control take too long for me.

(note any number i state above could easily be off a hundred Elo either way - Was never looking for greater precision, just ball park #'s.
I think the formulas are reversed. Where it says CCRL-Ra it should say FIDE-Ra and where it says FIDE-Ra it should say CCRL-Ra.

FIDE = (0.7 x CCRL) + 840
CCRL = (FIDE - 840) / 0.7

In my case I found that at the bottom of the list this was not well enforced. I now use FIDE = (0.9 x CCRL) + 440

Some time ago I tried to find an equivalence between the Basque Elo and the FIDE Elo, in principle there is not much difference. In the study I noticed an interesting thing, for example if I took players in the Basque list of 1700 points, I found in the FIDE list that these players could have Elos as 1900 points and Elos as 1100, an incredible difference. The average is something like 1511. By this I mean that no matter how well you calibrate your Elo you will always find people who will think that the level of play with the selected Elo is too high or too low even if you calibrate too much.

https://sites.google.com/site/danasah/c ... -elo-vasco (spanish)
When converting from FIDE to CCRL ratings or vice-versa, we need different formulae depending on which ratings we are talking about. FIDE has standard, Rapid, and Blitz ratings; CCRL has only Rapid and Blitz ratings. What the formulae should tell us is what CCRL Rapid rating a player with a FIDE Rapid rating of (for example) 2000 would get if he played against CCRL engines at their specified rapid time control and hardware, and (using a completely different formula) what CCRL Blitz rating a player with a FIDE Blitz rating of (for example) 2000 would get if he played against CCRL engines at their specified blitz time control (2' + 1") and hardware. Probably the data is only available for online blitz play between humans and engines, which I gather is at an average of something like 3' + 3" per move. So are these formulae based on data from such blitz games? If they are, then it would be very strange, really ridiculous, if they predicted that a 2800 FIDE human (even a blitz rating) would get a 2800 CCRL rating playing against CCRL engines in blitz, even if it is 3' + 3" rather than 2' + 1". So, just how are these formulae calibrated? Do they just "assume" that 2800 is the same on both scales? If so that is absurd for blitz.
Let's imagine that we want to organise a match between a GM and an engine and we want it to be evenly matched.

GM Elo = 2515 (standar)

With Kai's formula we could look for an engine from the CCRL list 40/15.

CCRL = (FIDE-840)/0.7 = 2392. In this Elo for example plays CT800 1.40 64-bit

The computer we have for the match is an i7-4770k, which matches the reference used by CCRL for the 40/15 time control. Let's consider for example a standard control something like 40/120.

What are the timings that would need to be set in the GUI for the GM and for the engine?

1. Human --> 40/120, Engine --> 40/120
or
2. Human --> 40/120, Engine --> 40/15
or
3. Human --> 40/15, Engine --> 40/15

Maybe the second option? In the second option the human has 8 times more time.

If we now play blitz for example 3'+3'', maybe the engine should be configured with 8 times less time? something like 22s + 375ms

Does this make any sense?

In the versions limited by strength for Elo FIDE you could divide the number of nps by 8 and use the same time for both.

I think this is pretty reasonable for the longer time limits at least. The CCRL ratings were probably intended to be about right at 40 moves in two hours based on the hardware of the first few years of this century, when the top humans and top engines (usually on 4 cpus) were pretty competitive. Roughly speaking, I think the reference hardware for CCRL is about four times as fast as the average hardware used in these events nearly twenty years ago. The ratings look about right for those old engines if you assume the hardware was 1/4 of the speed, maybe still a bit low. But if you play a time odds game, Ponder makes the true time odds significantly less than the nominal odds, so your 8 to 1 time odds proposal might in fact be about 4 to 1 when Ponder is considered, so it should indeed be about right for 40 moves in 2 hours for the human at 40 moves in 15 minutes for the engine if both are 2800, where the CCRL ratings and FIDE ratings are supposed to be comparable. But in blitz, the engine will still win with the 8 to 1 time odds, because human play degenerates much more than engine play with shorter time, although it's true that time odds weakens the engine more with shorter time. Maybe it would be reasonably competitive with Ponder off in blitz. I might try some games myself that way with a suitably rated engine. I'm only about 2200 level in blitz now, which would mean about a CCRL 1940 engine. Might be fair.

Ferdy · Post by **Ferdy** » Mon Apr 26, 2021 12:54 am

Ferdy wrote: ↑Tue Apr 20, 2021 12:02 pm The formula to get FIDE Elo from Lichess Blitz Elo is,
Code: Select all
FIDE Elo = 65 + [0.87 x Lichess Blitz Rating]

Update on Maia1 at Lichess blitz rating of 1510 is 1378 FIDE Elo.