UCI_Elo

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Fri Jul 12, 2019 6:21 pm

Danasah human is around 1300.

Code: Select all

   # PLAYER                              :  RATING  ERROR  POINTS  PLAYED   (%)
   1 Amyan 1.72 ucielo 1500              :  2351.6  137.5   112.0     132    85
   2 Cheese 2.1 ucielo 1500              :  2340.2  132.6   111.0     132    84
   3 Cheng 4.39 ucielo 1500              :  2329.1  132.5   110.0     132    83
   4 Fruit reloaded v3.21 ucielo 1500    :  2311.6  130.9   106.5     130    82
   5 Ufim v8.02 ucielo 1500              :  2146.3  118.6    99.5     146    68
   6 Rhetoric 1.4.3 ucielo 1500          :  2112.5  120.9    86.0     130    66
   7 DanaSah 7.9 ucielo 1500             :  2101.7  116.2    79.5     132    60
   8 MadChess 2.2 ucielo 1500            :  2088.8  115.2    92.0     146    63
   9 Houdini 3 ucielo 1500               :  2063.8  128.7    81.5     112    73
  10 D2019.2.37.53 ucielo 1500           :  2019.8  114.7    77.5     132    59
  11 Discocheck 5.2 ucielo 1500          :  1848.6  111.2    59.5     132    45
  12 Iota 1.0 ccrl 1019                  :  1821.1  158.6    15.5      46    34
  13 CT800 V1.34 ucielo 1500             :  1758.5  110.0    53.0     148    36
  14 Arasan 21.3 ucielo 1500             :  1662.1  112.8    41.5     132    31
  15 Hiarcs 14 ucielo 1500               :  1510.6  113.7    28.5     146    20
  16 NSVChess v0.14 ccrl 946             :  1500.0   ----    21.0     212    10
  17 DanaSah 7.9 human ucielo 1500       :  1278.1  156.1     7.5     208     4


User avatar
pedrox
Posts: 970
Joined: Fri Mar 10, 2006 5:07 am
Location: Basque Country (Spain)
Contact:

Re: UCI_Elo

Post by pedrox » Fri Jul 12, 2019 7:54 pm

Thanks for the results.

I had not seen anywhere that UCI_ELO refers to Elo FIDE. But it makes sense than when a user uses limit strenght is to play against the engine and in that case offer an Elo FIDE (Although I have also used my engine to deal with dedicated machines of the 80-90s). I will make the "human" version as the default version and I will study how to do the other options.

I think the "engine" version played more or less at the level I expected, the "human" version made it much lower than I expected. I will try to increase the strength for this mode by 200 Elo points.

In my engine, I could make a force adjustment by changing values in the configuration options. For example:

Diff engine = 50
Diff computer-engine = 350
Diff human-computer = 70

With these values it is possible that the engine in the "human" mode played something like 1500. But I will have to check if this is and if the regulation then works for other values.

Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Fri Jul 12, 2019 11:46 pm

According to uci protocol, UCI_Elo refers to Elo, since there is no other popular chess Elo than FIDE Elo, I believe this is FIDE Elo. Mark the author of Hiarcs is probably aware of this, his engine at 1500 uci elo is close.

Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Sat Jul 13, 2019 5:05 pm

A sample game against Danasah using the chess GUI that I have been developing, featuring 2 TC's one with time delay. Danasah played at TC 5min+10s (Fischer), and I am on TC 5min-10s (10s delay).

The pgn source has clk or clock showing the time remaining after a move. Press C8 on the board.




And a game with Arasan

MikeB
Posts: 3161
Joined: Thu Mar 09, 2006 5:34 am
Location: Pen Argyl, Pennsylvania

Re: UCI_Elo

Post by MikeB » Sat Jul 13, 2019 5:41 pm

pedrox wrote:
Fri Jul 12, 2019 7:54 pm
Thanks for the results.

I had not seen anywhere that UCI_ELO refers to Elo FIDE. But it makes sense than when a user uses limit strenght is to play against the engine and in that case offer an Elo FIDE (Although I have also used my engine to deal with dedicated machines of the 80-90s). I will make the "human" version as the default version and I will study how to do the other options.

I think the "engine" version played more or less at the level I expected, the "human" version made it much lower than I expected. I will try to increase the strength for this mode by 200 Elo points.

In my engine, I could make a force adjustment by changing values in the configuration options. For example:

Diff engine = 50
Diff computer-engine = 350
Diff human-computer = 70

With these values it is possible that the engine in the "human" mode played something like 1500. But I will have to check if this is and if the regulation then works for other values.
It is commonly accepted that Elo means something in the ball parl of FIDE Elo. Of course, many national federations have their own ratings systems and even the engine vs engine testers try have something that is supposed to align with FIDE Elo - but of course with no interaction between the universe of players between the human group of players and the engine universe of players, it is impossible to say CCRL equals FIDE, etc. We do know top players are rated around 2800 FIDE and it does appear from a distance , that an engine rated near 2800 CCRL is probably close to 2800 FIDE, but who really knows for sure. The answer is we do not know and , but we do know that is not exact - but it's probably in the range if you use very large bars - say 2800 CCRL is probably plus or minus 100 ELo of 2800 FIDE. And that of course will be true at 1500 ELO - plus or minus 100 Elo. And even my off the cuff comment here - somebody else will say , "no, it's xyz" and they could be right ..or they could be wrong. I would be shocked to find that the error bar would be more than 200 Elo off - but who knows. The very best players one the world no longer like to play the best engines in the world in public, and I I don't blame them one iota as the difference is now in the the multiple hundreds of ELO and they have almost no shot at winning even one game. Drawing a game now and then is probably the best they can do now.

Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Sun Jul 14, 2019 8:17 am

Reduce again Deuterium's nodes limit to 200, now it much closer to CT800. No other tricks added to reduce strenght so far, just the node reduction. Deuterium played on TC 3m+2s while I use TC 10m with 10s delay.

Code: Select all

   # PLAYER                                   :  RATING  ERROR  POINTS  PLAYED   (%)
   1 Amyan 1.72 ucielo 1500                   :  2298.3  118.2   120.0     140    86
   2 Cheng 4.39 ucielo 1500                   :  2281.8  117.3   118.5     140    85
   3 Cheese 2.1 ucielo 1500                   :  2245.1  118.0   115.0     140    82
   4 Fruit reloaded v3.21 ucielo 1500         :  2238.3  104.1   112.5     138    82
   5 Ufim v8.02 ucielo 1500                   :  2128.0   98.1   110.5     154    72
   6 Rhetoric 1.4.3 ucielo 1500               :  2062.9  108.5    93.5     138    68
   7 DanaSah 7.9 ucielo 1500                  :  2061.9  108.7    74.0     124    60
   8 MadChess 2.2 ucielo 1500                 :  2061.7   95.5   101.5     154    66
   9 Houdini 3 ucielo 1500                    :  2055.3  101.6    95.0     128    74
  10 Discocheck 5.2 ucielo 1500               :  1822.5  105.6    66.5     140    48
  11 Iota 1.0 ccrl 1019                       :  1822.1  121.7    24.5      58    42
  12 Deuterium v2019.2.37.53 ucielo 1500      :  1784.4   86.1    91.0     240    38
  13 CT800 V1.34 ucielo 1500                  :  1731.5   90.1    58.0     156    37
  14 Arasan 21.3 ucielo 1500                  :  1619.6   90.4    44.5     156    29
  15 Hiarcs 14 ucielo 1500                    :  1510.7  104.2    32.5     154    21
  16 NSVChess v0.14 ccrl 946                  :  1500.0   ----    25.0     212    12
  17 DanaSah 7.9 human ucielo 1500            :  1243.9  129.1     7.5     208     4

Sample game of how it played. It does not blunder material directly but you have to play with combination to outplay it.


Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Mon Jul 15, 2019 12:05 am

Reduce strength by randomizing piece values of queen and rook. Before a search is made, queen is randomize between 400 to 700 cp, while rook is between 300 and 700 cp. This is on Deuterium v2019.2.37.54 ucielo 1500. Note basic weakening is at 200 nodes a move at this 1500 Elo level plus this material randomizer.

The base version is Deuterium v2019.2.37.53 ucielo 1500 which is run at 200 nodes.

Result at TC 40/2min. It is now below CT800.

Code: Select all

   # PLAYER                                 :  RATING  ERROR  POINTS  PLAYED   (%)
   1 Amyan 1.72 ucielo 1500                 :  2290.3  118.7   135.5     156    87
   2 Cheng 4.39 ucielo 1500                 :  2279.5  131.5   134.5     156    86
   3 Cheese 2.1 ucielo 1500                 :  2234.1  118.4   130.0     156    83
   4 Fruit reloaded v3.21 ucielo 1500       :  2208.5  119.5   125.5     154    81
   5 Ufim v8.02 ucielo 1500                 :  2121.1  102.3   125.5     170    74
   6 MadChess 2.2 ucielo 1500               :  2065.4  105.2   117.5     170    69
   7 Rhetoric 1.4.3 ucielo 1500             :  2054.5  106.5   107.5     154    70
   8 DanaSah 7.9 ucielo 1500                :  2053.2  113.2    74.0     124    60
   9 Houdini 3 ucielo 1500                  :  2037.3  113.3   107.5     144    75
  10 Iota 1.0 ccrl 1019                     :  1821.9  131.7    24.5      58    42
  11 Discocheck 5.2 ucielo 1500             :  1819.8   96.7    77.5     156    50
  12 Deuterium v2019.2.37.53 ucielo 1500    :  1788.9   92.1   103.0     256    40
  13 CT800 V1.34 ucielo 1500                :  1727.9  108.0    67.0     172    39
  14 Deuterium v2019.2.37.54 ucielo 1500    :  1672.7   96.6    60.5     224    27
  15 Arasan 21.3 ucielo 1500                :  1628.3  102.4    52.5     172    31
  16 Hiarcs 14 ucielo 1500                  :  1511.7  108.3    37.0     170    22
  17 NSVChess v0.14 ccrl 946                :  1500.0   ----    25.0     212    12
  18 DanaSah 7.9 human ucielo 1500          :  1259.7  124.7     9.5     224     4

Sample game with Deuterium v2019.2.37.54 ucielo 1500. There is a bit of struggle in the opening and it keeps on weakening its squares. It gives up its queen without too much fight, perhaps this is because of the randomized queen value at [400, 700] cp.
I played at TC 10m and 10s delay, Deuterium is at TC 3m+2s.


Ferdy
Posts: 3938
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: UCI_Elo

Post by Ferdy » Fri Jul 19, 2019 11:37 am

Collected some games played by players with Elo 1400 to 1600, based from TWIC 2018 to July 2019. This could be useful to approximate the UCI_Elo 1500 for engine authors who may wish to implement UCI_Elo on their engine or improve current implementation.

The source pgn file is cleaned and doubles are removed by pgn-extract.

7000 plus games, white has a rating from 1400 to 1600.
https://drive.google.com/file/d/1eD0a9z ... sp=sharing

7000 plus games, black has a rating from 1400 to 1600.
https://drive.google.com/file/d/1eoXGxQ ... sp=sharing

Post Reply