UCI_Elo 1500 engines and FIDE Elo

[Modern Times]

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.

CCRL 40/40 ratings were set equal to a basket of engines on SSDF list at the time (2006), that list being 40/120 on an Athlon 1200 I think back then. That was simply so that there was a recognisable reference point rather then us thinking our time control and hardware combination was equal to theirs - although we thought it could be about the same, 40/120 on Athlon 1200 vs 40/40 an Athlon 64 3800+ . Subsequently we scaled the entire list down by 100 Elo. The top ratings on our list are very high, but personally I'm thankful we don't use Ordo which would expand the spread even more than bayeselo. I don't recall at the moment how we set the scale on the blitz list.

[MikeB]

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.

CCRL 40/40 ratings were set equal to a basket of engines on SSDF list at the time (2006), that list being 40/120 on an Athlon 1200 I think back then. That was simply so that there was a recognisable reference point rather then us thinking our time control and hardware combination was equal to theirs - although we thought it could be about the same, 40/120 on Athlon 1200 vs 40/40 an Athlon 64 3800+ . Subsequently we scaled the entire list down by 100 Elo. The top ratings on our list are very high, but personally I'm thankful we don't use Ordo which would expand the spread even more than bayeselo. I don't recall at the moment how we set the scale on the blitz list.

the answer is .. who really knows ? , but from everything I have seen, CCRL has taken a very reasonable approach ..and nobody really knows if some other way would have been better ...one artifact is that we get good scaling at the high-end, but the low-end is getting compressed (more compressed relative to human ratings) ...no rating system is perfect, that is one reason why we have so many ...every rating system I have seen , if they been around long enough , has had to make a rating adjustment of some sort ..FIDE, USCF, ICCF I think even SSDF has made rating adjustments if my memory serves me correctly ...

[Ferdy]

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.

Code: Select all

   # Player                                    :  FIDE Elo  Points  Played  Points(%)
   1 Equinox 3.30 UCI_Elo 1500                 :      2368   100.0     100      100.0
   2 Spike 1.4 UCI_Elo 1500                    :      2368   100.0     100      100.0
   3 Rybka v2.3.2a UCI_Elo 1500                :      2185    99.0     100       99.0
   4 Minic 2.51 UCI_Elo 1500                   :      2113    98.5     100       98.5
   5 Cheese 2.2 UCI_Elo 1500                   :      2113    98.5     100       98.5
   6 Cheese 2.1 UCI_Elo 1500                   :      2062    98.0     100       98.0
   7 Honey X5i UCI_Elo 1500                    :      2022    97.5     100       97.5
   8 Cheng 4.39 UCI_Elo 1500                   :      2022    97.5     100       97.5
   9 Ufim v8.02 UCI_Elo 1500                   :      1989    97.0     100       97.0
  10 Amyan 1.72 UCI_Elo 1500                   :      1878    94.5     100       94.5
  11 Wasp 4.5 UCI_Elo 1500                     :      1808    92.0     100       92.0
  12 MadChess 3.0 UCI_Elo 1500                 :      1785    91.0     100       91.0
  13 Deuterium v2021.1.38.29.4 UCI_Elo 1500    :      1721    87.5     100       87.5
  14 Rhetoric 1.4.3 UCI_Elo 1500               :      1705    86.5     100       86.5
  15 Discocheck 5.2 UCI_Elo 1500               :      1698    86.0     100       86.0
  16 Cheng 4.40 UCI_Elo 1500                   :      1698    86.0     100       86.0
  17 Arasan 21.1 UCI_Elo 1500                  :      1683    85.0     100       85.0
  18 Rodent IV 022 UCI_Elo 1500                :      1683    85.0     100       85.0
  19 Danasah 7.9 UCI_Elo 1500                  :      1567    74.5     100       74.5
  20 Deuterium v2019.2.37.73 UCI_Elo 1500      :      1567    74.5     100       74.5
  21 Stockfish 13 UCI_Elo 1500                 :      1507    67.5     100       67.5
  22 CT800 V1.42 UCI_Elo 1500                  :      1483    64.5     100       64.5
  23 Hiarcs 14 UCI_Elo 1500                    :      1417    55.5     100       55.5
  24 Maia1 Lichess 1510                        :      1378   363.5    2400       15.1
  25 Deuterium v2014.2.33.276 UCI_Elo 1500     :      1139    20.5     100       20.5

Ranking by distance from FIDE Elo 1500.

Code: Select all

 num                                  engine games  fide_elo  diff
   1               Stockfish 13 UCI_Elo 1500   100      1507     7
   2                CT800 V1.42 UCI_Elo 1500   100      1483    17
   3    Deuterium v2019.2.37.73 UCI_Elo 1500   100      1567    67
   4                Danasah 7.9 UCI_Elo 1500   100      1567    67
   5                  Hiarcs 14 UCI_Elo 1500   100      1417    83
   6                      Maia1 Lichess 1510  2400      1378   122
   7              Rodent IV 022 UCI_Elo 1500   100      1683   183
   8                Arasan 21.1 UCI_Elo 1500   100      1683   183
   9                 Cheng 4.40 UCI_Elo 1500   100      1698   198
  10             Discocheck 5.2 UCI_Elo 1500   100      1698   198
  11             Rhetoric 1.4.3 UCI_Elo 1500   100      1705   205
  12  Deuterium v2021.1.38.29.4 UCI_Elo 1500   100      1721   221
  13               MadChess 3.0 UCI_Elo 1500   100      1785   285
  14                   Wasp 4.5 UCI_Elo 1500   100      1808   308
  15   Deuterium v2014.2.33.276 UCI_Elo 1500   100      1139   361
  16                 Amyan 1.72 UCI_Elo 1500   100      1878   378
  17                 Ufim v8.02 UCI_Elo 1500   100      1989   489
  18                 Cheng 4.39 UCI_Elo 1500   100      2022   522
  19                  Honey X5i UCI_Elo 1500   100      2022   522
  20                 Cheese 2.1 UCI_Elo 1500   100      2062   562
  21                 Cheese 2.2 UCI_Elo 1500   100      2113   613
  22                 Minic 2.51 UCI_Elo 1500   100      2113   613
  23              Rybka v2.3.2a UCI_Elo 1500   100      2185   685
  24                  Spike 1.4 UCI_Elo 1500   100      2368   868
  25               Equinox 3.30 UCI_Elo 1500   100      2368   868
  
  diff : The absolute difference between 1500 and fide_elo, lower diff is better.

Do you have enough games with this setting against unrestricted (weak) engines with CCRL blitz ratings at 2' + 1" to estimate what CCRL blitz rating it would get (at 2' + 1")? This would tell us the conversion between CCRL and FIDE at that level anyhow.

I have not done such test yet, perhaps next week.

Am I interpreting the above correctly if I say that based on the best evidence, Stockfish 13 set to play at elo 1500 will perform about evenly with humans FIDE rated around 1500 (1507 per the above chart) if they are both playing blitz at a time control averaging something like 3' + 3"?

The data in the regression in 2019 was Lichess Blitz and FIDE Rating as in the image.



I believe the FIDE Rating in the image is the std rating. So this can be:

Code: Select all

FIDE Std. Elo = 65 + [0.87 x Lichess Blitz Rating]

Based on some games I ran with the ELO 1500 to 1700 range of SF13 vs "Irina 0.15" (CCRL blitz 1436), I calculate that you would subtract 224 elo from the nominal SF ELO to estimate its CCRL blitz rating. So if I am interpreting everything properly, and based on your indications that the SF elo 1500 level would indeed be a close match playing lichess blitz with humans rated about 1500 (1507) on the FIDE scale, this would indicate that a CCRL engine with a blitz rating of 1276 would be an even match playing blitz with a human rated FIDE 1500 (or 1507). So "Irina 0.15" should be a fair match in blitz for a human rated 1660 FIDE (or perhaps a bit less as the FIDE scale is more compact than CCRL). This doesn't seem plausible to me, humans of that level playing blitz really blunder a lot of pieces, but perhaps I'm wrong, or perhaps I'm misunderstanding something. Could it be that engines playing on Lichess are much weaker than the same engine/settings playing on a fast i7 (single thread)?

The reference engine maia1 that I use is run on Lc0 node=1 limit, the same setting is used in Lichess, so this is easy to reproduce. However the ranking may change if I will change the TC from 60s+50ms bullet to TC 180s+1s for example depends on the engines' weakening algorithm. Have not tested it though.

[lkaufman]

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.

Code: Select all

   # Player                                    :  FIDE Elo  Points  Played  Points(%)
   1 Equinox 3.30 UCI_Elo 1500                 :      2368   100.0     100      100.0
   2 Spike 1.4 UCI_Elo 1500                    :      2368   100.0     100      100.0
   3 Rybka v2.3.2a UCI_Elo 1500                :      2185    99.0     100       99.0
   4 Minic 2.51 UCI_Elo 1500                   :      2113    98.5     100       98.5
   5 Cheese 2.2 UCI_Elo 1500                   :      2113    98.5     100       98.5
   6 Cheese 2.1 UCI_Elo 1500                   :      2062    98.0     100       98.0
   7 Honey X5i UCI_Elo 1500                    :      2022    97.5     100       97.5
   8 Cheng 4.39 UCI_Elo 1500                   :      2022    97.5     100       97.5
   9 Ufim v8.02 UCI_Elo 1500                   :      1989    97.0     100       97.0
  10 Amyan 1.72 UCI_Elo 1500                   :      1878    94.5     100       94.5
  11 Wasp 4.5 UCI_Elo 1500                     :      1808    92.0     100       92.0
  12 MadChess 3.0 UCI_Elo 1500                 :      1785    91.0     100       91.0
  13 Deuterium v2021.1.38.29.4 UCI_Elo 1500    :      1721    87.5     100       87.5
  14 Rhetoric 1.4.3 UCI_Elo 1500               :      1705    86.5     100       86.5
  15 Discocheck 5.2 UCI_Elo 1500               :      1698    86.0     100       86.0
  16 Cheng 4.40 UCI_Elo 1500                   :      1698    86.0     100       86.0
  17 Arasan 21.1 UCI_Elo 1500                  :      1683    85.0     100       85.0
  18 Rodent IV 022 UCI_Elo 1500                :      1683    85.0     100       85.0
  19 Danasah 7.9 UCI_Elo 1500                  :      1567    74.5     100       74.5
  20 Deuterium v2019.2.37.73 UCI_Elo 1500      :      1567    74.5     100       74.5
  21 Stockfish 13 UCI_Elo 1500                 :      1507    67.5     100       67.5
  22 CT800 V1.42 UCI_Elo 1500                  :      1483    64.5     100       64.5
  23 Hiarcs 14 UCI_Elo 1500                    :      1417    55.5     100       55.5
  24 Maia1 Lichess 1510                        :      1378   363.5    2400       15.1
  25 Deuterium v2014.2.33.276 UCI_Elo 1500     :      1139    20.5     100       20.5

Ranking by distance from FIDE Elo 1500.

Code: Select all

 num                                  engine games  fide_elo  diff
   1               Stockfish 13 UCI_Elo 1500   100      1507     7
   2                CT800 V1.42 UCI_Elo 1500   100      1483    17
   3    Deuterium v2019.2.37.73 UCI_Elo 1500   100      1567    67
   4                Danasah 7.9 UCI_Elo 1500   100      1567    67
   5                  Hiarcs 14 UCI_Elo 1500   100      1417    83
   6                      Maia1 Lichess 1510  2400      1378   122
   7              Rodent IV 022 UCI_Elo 1500   100      1683   183
   8                Arasan 21.1 UCI_Elo 1500   100      1683   183
   9                 Cheng 4.40 UCI_Elo 1500   100      1698   198
  10             Discocheck 5.2 UCI_Elo 1500   100      1698   198
  11             Rhetoric 1.4.3 UCI_Elo 1500   100      1705   205
  12  Deuterium v2021.1.38.29.4 UCI_Elo 1500   100      1721   221
  13               MadChess 3.0 UCI_Elo 1500   100      1785   285
  14                   Wasp 4.5 UCI_Elo 1500   100      1808   308
  15   Deuterium v2014.2.33.276 UCI_Elo 1500   100      1139   361
  16                 Amyan 1.72 UCI_Elo 1500   100      1878   378
  17                 Ufim v8.02 UCI_Elo 1500   100      1989   489
  18                 Cheng 4.39 UCI_Elo 1500   100      2022   522
  19                  Honey X5i UCI_Elo 1500   100      2022   522
  20                 Cheese 2.1 UCI_Elo 1500   100      2062   562
  21                 Cheese 2.2 UCI_Elo 1500   100      2113   613
  22                 Minic 2.51 UCI_Elo 1500   100      2113   613
  23              Rybka v2.3.2a UCI_Elo 1500   100      2185   685
  24                  Spike 1.4 UCI_Elo 1500   100      2368   868
  25               Equinox 3.30 UCI_Elo 1500   100      2368   868
  
  diff : The absolute difference between 1500 and fide_elo, lower diff is better.

Do you have enough games with this setting against unrestricted (weak) engines with CCRL blitz ratings at 2' + 1" to estimate what CCRL blitz rating it would get (at 2' + 1")? This would tell us the conversion between CCRL and FIDE at that level anyhow.

I have not done such test yet, perhaps next week.

Am I interpreting the above correctly if I say that based on the best evidence, Stockfish 13 set to play at elo 1500 will perform about evenly with humans FIDE rated around 1500 (1507 per the above chart) if they are both playing blitz at a time control averaging something like 3' + 3"?

The data in the regression in 2019 was Lichess Blitz and FIDE Rating as in the image.



I believe the FIDE Rating in the image is the std rating. So this can be:

Code: Select all

FIDE Std. Elo = 65 + [0.87 x Lichess Blitz Rating]

Based on some games I ran with the ELO 1500 to 1700 range of SF13 vs "Irina 0.15" (CCRL blitz 1436), I calculate that you would subtract 224 elo from the nominal SF ELO to estimate its CCRL blitz rating. So if I am interpreting everything properly, and based on your indications that the SF elo 1500 level would indeed be a close match playing lichess blitz with humans rated about 1500 (1507) on the FIDE scale, this would indicate that a CCRL engine with a blitz rating of 1276 would be an even match playing blitz with a human rated FIDE 1500 (or 1507). So "Irina 0.15" should be a fair match in blitz for a human rated 1660 FIDE (or perhaps a bit less as the FIDE scale is more compact than CCRL). This doesn't seem plausible to me, humans of that level playing blitz really blunder a lot of pieces, but perhaps I'm wrong, or perhaps I'm misunderstanding something. Could it be that engines playing on Lichess are much weaker than the same engine/settings playing on a fast i7 (single thread)?

The reference engine maia1 that I use is run on Lc0 node=1 limit, the same setting is used in Lichess, so this is easy to reproduce. However the ranking may change if I will change the TC from 60s+50ms bullet to TC 180s+1s for example depends on the engines' weakening algorithm. Have not tested it though.

OK, so this is at least a partial explanation; presumably the crippled SF 1500 would play better with more time (even an engine with bad eval like the crippled SF should improve with time), so if the Lichess games were typically played with something like quadruple the total time (for 60 moves), it would be likely that SF "1500" might get a rating there more like 1650 or 1700. So "Irina 0.15" might be more like an even match for a FIDE 1800 or a bit higher in blitz based on your data. This is at least within the realm of the plausible to me, although I would guess that a FIDE rating of 1900 to 2000 would be needed for this. I wonder what the typical time limit is for these Lichess games; although humans prefer fast blitz like 3' no inc. with each other, they may choose relatively long time limits/increments when playing engines. But I don't play on Lichess so I have to ask others to comment on this.

[KLc]

The reference engine maia1 that I use is run on Lc0 node=1 limit, the same setting is used in Lichess, so this is easy to reproduce. However the ranking may change if I will change the TC from 60s+50ms bullet to TC 180s+1s for example depends on the engines' weakening algorithm. Have not tested it though.

Exactly, it‘s fixed to nodes=1. But doesn’t this mean it plays independently of time control? The only difference can be how humans perform on different time controls against the exact same opponent. No?

[Ferdy]

The reference engine maia1 that I use is run on Lc0 node=1 limit, the same setting is used in Lichess, so this is easy to reproduce. However the ranking may change if I will change the TC from 60s+50ms bullet to TC 180s+1s for example depends on the engines' weakening algorithm. Have not tested it though.

Exactly, it‘s fixed to nodes=1. But doesn’t this mean it plays independently of time control? The only difference can be how humans perform on different time controls against the exact same opponent. No?

Yes there is no time control in maia1's setting, it is replaced by node control, the idea is that it can play on different hardware and get the same strength every time, the time element is erased. Human strength can vary due to time but perhaps not much at this Elo 1500 level.

Based from regression,
FIDE Std. Elo = 65 + 0.87 x Lichess Blitz Rating

Lichess Blitz Rating = (FIDE Std. Elo - 65) / 0.87

FIDE Std. Elo = 1500
Lichess Blitz Rating = (1500 - 65) / 0.87 = 1650

We can extract data from people with Lichess blitz rating of around 1650, then get its rapid rating, do regression and see how close these people can get with added time that is from blitz to rapid.

[Ferdy]

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.

Code: Select all

   # Player                                    :  FIDE Elo  Points  Played  Points(%)
   1 Equinox 3.30 UCI_Elo 1500                 :      2368   100.0     100      100.0
   2 Spike 1.4 UCI_Elo 1500                    :      2368   100.0     100      100.0
   3 Rybka v2.3.2a UCI_Elo 1500                :      2185    99.0     100       99.0
   4 Minic 2.51 UCI_Elo 1500                   :      2113    98.5     100       98.5
   5 Cheese 2.2 UCI_Elo 1500                   :      2113    98.5     100       98.5
   6 Cheese 2.1 UCI_Elo 1500                   :      2062    98.0     100       98.0
   7 Honey X5i UCI_Elo 1500                    :      2022    97.5     100       97.5
   8 Cheng 4.39 UCI_Elo 1500                   :      2022    97.5     100       97.5
   9 Ufim v8.02 UCI_Elo 1500                   :      1989    97.0     100       97.0
  10 Amyan 1.72 UCI_Elo 1500                   :      1878    94.5     100       94.5
  11 Wasp 4.5 UCI_Elo 1500                     :      1808    92.0     100       92.0
  12 MadChess 3.0 UCI_Elo 1500                 :      1785    91.0     100       91.0
  13 Deuterium v2021.1.38.29.4 UCI_Elo 1500    :      1721    87.5     100       87.5
  14 Rhetoric 1.4.3 UCI_Elo 1500               :      1705    86.5     100       86.5
  15 Discocheck 5.2 UCI_Elo 1500               :      1698    86.0     100       86.0
  16 Cheng 4.40 UCI_Elo 1500                   :      1698    86.0     100       86.0
  17 Arasan 21.1 UCI_Elo 1500                  :      1683    85.0     100       85.0
  18 Rodent IV 022 UCI_Elo 1500                :      1683    85.0     100       85.0
  19 Danasah 7.9 UCI_Elo 1500                  :      1567    74.5     100       74.5
  20 Deuterium v2019.2.37.73 UCI_Elo 1500      :      1567    74.5     100       74.5
  21 Stockfish 13 UCI_Elo 1500                 :      1507    67.5     100       67.5
  22 CT800 V1.42 UCI_Elo 1500                  :      1483    64.5     100       64.5
  23 Hiarcs 14 UCI_Elo 1500                    :      1417    55.5     100       55.5
  24 Maia1 Lichess 1510                        :      1378   363.5    2400       15.1
  25 Deuterium v2014.2.33.276 UCI_Elo 1500     :      1139    20.5     100       20.5

Ranking by distance from FIDE Elo 1500.

Code: Select all

 num                                  engine games  fide_elo  diff
   1               Stockfish 13 UCI_Elo 1500   100      1507     7
   2                CT800 V1.42 UCI_Elo 1500   100      1483    17
   3    Deuterium v2019.2.37.73 UCI_Elo 1500   100      1567    67
   4                Danasah 7.9 UCI_Elo 1500   100      1567    67
   5                  Hiarcs 14 UCI_Elo 1500   100      1417    83
   6                      Maia1 Lichess 1510  2400      1378   122
   7              Rodent IV 022 UCI_Elo 1500   100      1683   183
   8                Arasan 21.1 UCI_Elo 1500   100      1683   183
   9                 Cheng 4.40 UCI_Elo 1500   100      1698   198
  10             Discocheck 5.2 UCI_Elo 1500   100      1698   198
  11             Rhetoric 1.4.3 UCI_Elo 1500   100      1705   205
  12  Deuterium v2021.1.38.29.4 UCI_Elo 1500   100      1721   221
  13               MadChess 3.0 UCI_Elo 1500   100      1785   285
  14                   Wasp 4.5 UCI_Elo 1500   100      1808   308
  15   Deuterium v2014.2.33.276 UCI_Elo 1500   100      1139   361
  16                 Amyan 1.72 UCI_Elo 1500   100      1878   378
  17                 Ufim v8.02 UCI_Elo 1500   100      1989   489
  18                 Cheng 4.39 UCI_Elo 1500   100      2022   522
  19                  Honey X5i UCI_Elo 1500   100      2022   522
  20                 Cheese 2.1 UCI_Elo 1500   100      2062   562
  21                 Cheese 2.2 UCI_Elo 1500   100      2113   613
  22                 Minic 2.51 UCI_Elo 1500   100      2113   613
  23              Rybka v2.3.2a UCI_Elo 1500   100      2185   685
  24                  Spike 1.4 UCI_Elo 1500   100      2368   868
  25               Equinox 3.30 UCI_Elo 1500   100      2368   868
  
  diff : The absolute difference between 1500 and fide_elo, lower diff is better.

Do you have enough games with this setting against unrestricted (weak) engines with CCRL blitz ratings at 2' + 1" to estimate what CCRL blitz rating it would get (at 2' + 1")? This would tell us the conversion between CCRL and FIDE at that level anyhow.

I have not done such test yet, perhaps next week.

Am I interpreting the above correctly if I say that based on the best evidence, Stockfish 13 set to play at elo 1500 will perform about evenly with humans FIDE rated around 1500 (1507 per the above chart) if they are both playing blitz at a time control averaging something like 3' + 3"?

The data in the regression in 2019 was Lichess Blitz and FIDE Rating as in the image.



I believe the FIDE Rating in the image is the std rating. So this can be:

Code: Select all

FIDE Std. Elo = 65 + [0.87 x Lichess Blitz Rating]

Based on some games I ran with the ELO 1500 to 1700 range of SF13 vs "Irina 0.15" (CCRL blitz 1436), I calculate that you would subtract 224 elo from the nominal SF ELO to estimate its CCRL blitz rating. So if I am interpreting everything properly, and based on your indications that the SF elo 1500 level would indeed be a close match playing lichess blitz with humans rated about 1500 (1507) on the FIDE scale, this would indicate that a CCRL engine with a blitz rating of 1276 would be an even match playing blitz with a human rated FIDE 1500 (or 1507). So "Irina 0.15" should be a fair match in blitz for a human rated 1660 FIDE (or perhaps a bit less as the FIDE scale is more compact than CCRL). This doesn't seem plausible to me, humans of that level playing blitz really blunder a lot of pieces, but perhaps I'm wrong, or perhaps I'm misunderstanding something. Could it be that engines playing on Lichess are much weaker than the same engine/settings playing on a fast i7 (single thread)?

The reference engine maia1 that I use is run on Lc0 node=1 limit, the same setting is used in Lichess, so this is easy to reproduce. However the ranking may change if I will change the TC from 60s+50ms bullet to TC 180s+1s for example depends on the engines' weakening algorithm. Have not tested it though.

OK, so this is at least a partial explanation; presumably the crippled SF 1500 would play better with more time (even an engine with bad eval like the crippled SF should improve with time), so if the Lichess games were typically played with something like quadruple the total time (for 60 moves), it would be likely that SF "1500" might get a rating there more like 1650 or 1700. So "Irina 0.15" might be more like an even match for a FIDE 1800 or a bit higher in blitz based on your data. This is at least within the realm of the plausible to me, although I would guess that a FIDE rating of 1900 to 2000 would be needed for this. I wonder what the typical time limit is for these Lichess games; although humans prefer fast blitz like 3' no inc. with each other, they may choose relatively long time limits/increments when playing engines. But I don't play on Lichess so I have to ask others to comment on this.

Tested the engines at blitz tc on my pc at 3m+1s around 2m+1s in ccrl blitz.

Blitz 3m+1s

Code: Select all

   # Player                                   :  FIDE Elo  Points  Played  Points(%)
   1 Honey X5i UCI_Elo 1500                   :      2370   100.0     100      100.0
   2 Amyan 1.72 UCI_Elo 1500                  :      2370   100.0     100      100.0
   3 Equinox 3.30 UCI_Elo 1500                :      2307    99.5     100       99.5
   4 Spike 1.4 UCI_Elo 1500                   :      2185    99.0     100       99.0
   5 Cheese 2.2 UCI_Elo 1500                  :      2113    98.5     100       98.5
   6 Cheng 4.40 UCI_Elo 1500                  :      2062    98.0     100       98.0
   7 Ufim v8.02 UCI_Elo 1500                  :      1989    97.0     100       97.0
   8 Danasah 7.9 UCI_Elo 1500                 :      1961    96.5     100       96.5
   9 Rybka v2.3.2a UCI_Elo 1500               :      1937    96.0     100       96.0
  10 Rodent IV 022 UCI_Elo 1500               :      1937    96.0     100       96.0
  11 Rhetoric 1.4.3 UCI_Elo 1500              :      1937    96.0     100       96.0
  12 Discocheck 5.2 UCI_Elo 1500              :      1862    94.0     100       94.0
  13 Minic 2.51 UCI_Elo 1500                  :      1847    93.5     100       93.5
  14 Wasp 4.5 UCI_Elo 1500                    :      1820    92.5     100       92.5
  15 CT800 V1.42 UCI_Elo 1500                 :      1764    90.0     100       90.0
  16 MadChess 3.0 UCI_Elo 1500                :      1720    87.5     100       87.5
  17 Arasan 21.1 UCI_Elo 1500                 :      1683    85.0     100       85.0
  18 Stockfish 13 UCI_Elo 1500                :      1527    70.0     100       70.0
  19 Deuterium v2019.2.37.73 UCI_Elo 1500     :      1519    69.0     100       69.0
  20 Hiarcs 14 UCI_Elo 1500                   :      1487    65.0     100       65.0
  21 Maia1 Lichess Blitz 1510                 :      1378   242.5    2100       11.5
  22 Deuterium v2014.2.33.276 UCI_Elo 1500    :      1265    34.5     100       34.5

Bullet 60s+100ms

Code: Select all

  # Player                                    :  FIDE Elo  Points  Played  Points(%)
   1 Equinox 3.30 UCI_Elo 1500                 :      2368   100.0     100      100.0
   2 Spike 1.4 UCI_Elo 1500                    :      2368   100.0     100      100.0
   3 Rybka v2.3.2a UCI_Elo 1500                :      2185    99.0     100       99.0
   4 Minic 2.51 UCI_Elo 1500                   :      2113    98.5     100       98.5
   5 Cheese 2.2 UCI_Elo 1500                   :      2113    98.5     100       98.5
   6 Cheese 2.1 UCI_Elo 1500                   :      2062    98.0     100       98.0
   7 Honey X5i UCI_Elo 1500                    :      2022    97.5     100       97.5
   8 Cheng 4.39 UCI_Elo 1500                   :      2022    97.5     100       97.5
   9 Ufim v8.02 UCI_Elo 1500                   :      1989    97.0     100       97.0
  10 Amyan 1.72 UCI_Elo 1500                   :      1878    94.5     100       94.5
  11 Wasp 4.5 UCI_Elo 1500                     :      1808    92.0     100       92.0
  12 MadChess 3.0 UCI_Elo 1500                 :      1785    91.0     100       91.0
  13 Deuterium v2021.1.38.29.4 UCI_Elo 1500    :      1721    87.5     100       87.5
  14 Rhetoric 1.4.3 UCI_Elo 1500               :      1705    86.5     100       86.5
  15 Discocheck 5.2 UCI_Elo 1500               :      1698    86.0     100       86.0
  16 Cheng 4.40 UCI_Elo 1500                   :      1698    86.0     100       86.0
  17 Arasan 21.1 UCI_Elo 1500                  :      1683    85.0     100       85.0
  18 Rodent IV 022 UCI_Elo 1500                :      1683    85.0     100       85.0
  19 Danasah 7.9 UCI_Elo 1500                  :      1567    74.5     100       74.5
  20 Deuterium v2019.2.37.73 UCI_Elo 1500      :      1567    74.5     100       74.5
  21 Stockfish 13 UCI_Elo 1500                 :      1507    67.5     100       67.5
  22 CT800 V1.42 UCI_Elo 1500                  :      1483    64.5     100       64.5
  23 Hiarcs 14 UCI_Elo 1500                    :      1417    55.5     100       55.5
  24 Maia1 Lichess Blitz 1510                  :      1378   363.5    2400       15.1
  25 Deuterium v2014.2.33.276 UCI_Elo 1500     :      1139    20.5     100       20.5

Stockfish, Hiarcs and Deuterium v2019.2.37.73 are still close to FIDE Elo 1500.

[Patrice Duhamel]

Stockfish, Hiarcs and Deuterium v2019.2.37.73 are still close to FIDE Elo 1500.

On my computer (i5 2500k) when I run 100 games at 60s+100ms with Cheese 2.2 vs Deuterium 2.37.73, both at UCI_Elo = 1500, Deuterium is +188 ELO better than Cheese.

Am I doing something wrong ?

[Ferdy]

Stockfish, Hiarcs and Deuterium v2019.2.37.73 are still close to FIDE Elo 1500.

On my computer (i5 2500k) when I run 100 games at 60s+100ms with Cheese 2.2 vs Deuterium 2.37.73, both at UCI_Elo = 1500, Deuterium is +188 ELO better than Cheese.

Am I doing something wrong ?

Try Cheese 2.2 vs Maia1, use nodes=1 for Maia lets see the result.

Later I will try Cheese 2.2 vs Deuterium.

[Ferdy]

This is my result for Cheese vs Deuterium.

Code: Select all

Score of Cheese 2.2 UCI_Elo 1500 vs Deuterium v2019.2.37.73 UCI_Elo 1500: 29 - 0 - 0  [1.000] 29
...      Cheese 2.2 UCI_Elo 1500 playing White: 15 - 0 - 0  [1.000] 15
...      Cheese 2.2 UCI_Elo 1500 playing Black: 14 - 0 - 0  [1.000] 14
...      White vs Black: 15 - 14 - 0  [0.517] 29
Elo difference: inf +/- nan, LOS: 100.0 %, DrawRatio: 0.0 %

I can see in the log that cutechess-cli has sent the proper commands to Cheese.

1483 <Cheese 2.2 UCI_Elo 1500(0): Cheese 2.2 (64 bits) by Patrice Duhamel
1484 <Cheese 2.2 UCI_Elo 1500(0): id name Cheese 2.2 64 bits
1484 <Cheese 2.2 UCI_Elo 1500(0): id author Patrice Duhamel
1484 <Cheese 2.2 UCI_Elo 1500(0): option name Hash type spin default 512 min 1 max 16384
1484 <Cheese 2.2 UCI_Elo 1500(0): option name Ponder type check default false
1484 <Cheese 2.2 UCI_Elo 1500(0): option name OwnBook type check default false
1485 <Cheese 2.2 UCI_Elo 1500(0): option name Clear Hash type button
1485 <Cheese 2.2 UCI_Elo 1500(0): option name UsePersonality type check default false
1485 <Cheese 2.2 UCI_Elo 1500(0): option name PersonalityFile type string default personality/default.txt
1486 <Cheese 2.2 UCI_Elo 1500(0): option name UCI_Chess960 type check default false
1486 <Cheese 2.2 UCI_Elo 1500(0): option name MultiPV type spin default 1 min 1 max 32
1486 <Cheese 2.2 UCI_Elo 1500(0): option name NullMovePruning type check default true
1486 <Cheese 2.2 UCI_Elo 1500(0): option name LateMoveReduction type check default true
1487 <Cheese 2.2 UCI_Elo 1500(0): option name Threads type spin default 1 min 1 max 64
1487 <Cheese 2.2 UCI_Elo 1500(0): option name UCI_LimitStrength type check default false
1487 <Cheese 2.2 UCI_Elo 1500(0): option name UCI_Elo type spin default 3200 min 1000 max 3200
1488 <Cheese 2.2 UCI_Elo 1500(0): option name DrawScore type spin default 0 min -100 max 100
1488 <Cheese 2.2 UCI_Elo 1500(0): uciok
1488 >Cheese 2.2 UCI_Elo 1500(0): setoption name DrawScore value 0
1488 >Cheese 2.2 UCI_Elo 1500(0): setoption name Hash value 128
1488 >Cheese 2.2 UCI_Elo 1500(0): setoption name LateMoveReduction value true
1489 >Cheese 2.2 UCI_Elo 1500(0): setoption name MultiPV value 1
1489 >Cheese 2.2 UCI_Elo 1500(0): setoption name NullMovePruning value true
1489 >Cheese 2.2 UCI_Elo 1500(0): setoption name OwnBook value false
1489 >Cheese 2.2 UCI_Elo 1500(0): setoption name PersonalityFile value personality/default.txt
1490 >Cheese 2.2 UCI_Elo 1500(0): setoption name Threads value 1
1490 >Cheese 2.2 UCI_Elo 1500(0): setoption name UCI_Elo value 1500
1490 >Cheese 2.2 UCI_Elo 1500(0): setoption name UCI_LimitStrength value true
1490 >Cheese 2.2 UCI_Elo 1500(0): setoption name UsePersonality value false

And also to Deuterium.

1655 >Deuterium v2019.2.37.73 UCI_Elo 1500(1): setoption name UCI_Elo value 1500
1655 >Deuterium v2019.2.37.73 UCI_Elo 1500(1): setoption name UCI_LimitStrength value true

Sample game:

[Event "?"]
[Site "?"]
[Date "2021.05.07"]
[Round "1"]
[White "Cheese 2.2 UCI_Elo 1500"]
[Black "Deuterium v2019.2.37.73 UCI_Elo 1500"]
[Result "1-0"]
[ECO "A50"]
[GameDuration "00:00:48"]
[GameEndTime "2021-05-07T01:00:31.440 Taipei Standard Time"]
[GameStartTime "2021-05-07T00:59:42.517 Taipei Standard Time"]
[Opening "Queen's pawn game"]
[PlyCount "87"]
[TimeControl "60+0.1"]

1. d4 {book} Nf6 {book} 2. c4 {book} a6 {book} 3. Nf3 {book} d5 {-0.67/4 0.008s}
4. cxd5 {+0.73/8 4.24s} Qxd5 {+0.73/3 0.003s} 5. Nc3 {+0.78/8 3.75s}
Qf5 {-2.92/2 0.003s} 6. e3 {+1.18/7 3.32s} Nc6 {-1.33/3 0.003s}
7. Bd3 {+1.02/8 2.76s} Qd7 {-2.61/2 0.003s} 8. O-O {+1.34/7 1.60s}
e6 {-1.60/3 0.001s} 9. Re1 {+1.13/6 1.60s} Bb4 {-4.58/3 0.003s}
10. a3 {+0.93/7 2.18s} Bxc3 {-2.28/4 0.005s} 11. bxc3 {+0.92/7 1.93s}
Qd6 {-2.10/3 0.003s} 12. a4 {+1.69/6 1.14s} O-O {-2.33/3 0.003s}
13. Ba3 {+2.18/8 1.53s} Qd8 {-1.83/3 0.004s} 14. Bxf8 {+2.15/7 1.35s}
Qxf8 {-6.94/3 0.004s} 15. Re2 {+2.26/6 0.80s} Bd7 {-4.27/4 0.004s}
16. Rb2 {+2.23/7 1.28s} b6 {-5.05/3 0.004s} 17. Ng5 {+2.23/6 0.89s}
h6 {-2.80/3 0.003s} 18. Ne4 {+2.23/7 0.86s} Nh7 {-1.49/3 0.003s}
19. Qe2 {+2.87/6 1.20s} f5 {+0.41/3 0.003s} 20. Ng3 {+2.67/6 0.82s}
Nf6 {+0.07/3 0.003s} 21. Bxa6 {+2.70/6 0.95s} Nb8 {-0.56/3 0.003s}
22. Bc4 {+3.48/6 1.11s} Ra5 {-0.83/3 0.002s} 23. Qf3 {+3.47/5 0.73s}
Nc6 {-3.88/3 0.004s} 24. Nxf5 {+4.12/5 0.83s} Kh8 {-2.28/2 0.002s}
25. Ng3 {+4.17/6 0.92s} g5 {-3.02/2 0.003s} 26. e4 {+4.84/5 0.92s}
g4 {-0.47/3 0.003s} 27. Qf4 {+4.82/7 0.96s} e5 {-6.25/3 0.003s}
28. Qe3 {+4.78/6 0.95s} Ra8 {-7.06/2 0.003s} 29. Qd3 {+4.71/6 0.92s}
exd4 {-7.34/2 0.001s} 30. cxd4 {+4.95/6 0.62s} Nb4 {-4.30/3 0.003s}
31. Qc3 {+5.11/5 0.86s} Nxe4 {-6.20/2 0.003s} 32. Nxe4 {+7.85/4 0.83s}
Re8 {-7.31/2 0.002s} 33. Nc5 {+7.81/5 0.83s} bxc5 {-11.82/2 0.002s}
34. dxc5+ {+7.78/6 0.65s} Qg7 {-12.34/5 0.004s} 35. Qxb4 {+7.89/7 0.74s}
Bc6 {-12.10/3 0.003s} 36. Rd1 {+7.93/5 0.76s} Bf3 {-7.57/2 0.003s}
37. Rd3 {+7.87/4 0.55s} Qe5 {-3.48/3 0.002s} 38. gxf3 {+11.42/6 0.61s}
gxf3 {-4.73/3 0.003s} 39. Rxf3 {+11.49/5 0.70s} c6 {+1.59/1 0.002s}
40. Rg3 {+11.72/5 0.67s} Qd4 {-14.22/2 0.002s} 41. Qb7 {+12.97/4 0.58s}
Qxb2 {-12.31/1 0.003s} 42. Qxb2+ {+M5/6 0.40s} Re5 {-M4/3 0.003s}
43. Qxe5+ {+M3/6 0.18s} Kh7 {-M2/2 0.001s} 44. Bd3# {+M1/6 0.058s, White mates}
1-0

Cheese searches deeper, Deuterium is only around depth 3.