Fun with foreign Piece values and PST's

[Rebel]

http://www.top-5000.nl/prodeo187x.htm

[lucasart]

http://www.top-5000.nl/prodeo187x.htm

Cool stuff! I wonder if modifying the PST and material values is enough to be classified as a potential clone by the dendogram people. It would be interesting to test ProDeo with personality X for correlation to engine X, and see what comes out of it.

If the conclusion is positive, then it demonstrates that the statistical detection of clones is dubious: an honest programmer who writes the entire engine and borrows PST is considered a "cheat", while the real "cheat" will be clever enough to change the PST and come out clean

Perhaps I will put my PST in an external binary file. That way, and if we can agree on the binary format for PST+Material, our engines can load any kind of "personality", and people can contribute personality files, etc. Also, perhaps bishop pair should be part of the material values.

[velmarin]

Engines come to mind as ZCT by Zach Wegner, or Bison,
with external files to update eval.

A user-modifiable Prodeo would be appreciated by many fans, loving personalities.

[Tennison]

Good idea.

It's easy to verify : Testing Prodeo (Adam Hair) and Toga Hair with clones detection Tools ...

[Adam Hair]

http://www.top-5000.nl/prodeo187x.htm

Cool stuff! I wonder if modifying the PST and material values is enough to be classified as a potential clone by the dendogram people. It would be interesting to test ProDeo with personality X for correlation to engine X, and see what comes out of it.

If the conclusion is positive, then it demonstrates that the statistical detection of clones is dubious: an honest programmer who writes the entire engine and borrows PST is considered a "cheat", while the real "cheat" will be clever enough to change the PST and come out clean

It has been tried. Swapping pst values did not change the similarity percentages enough to make a difference. However, the engines used have advance evaluation functions and their psts are relatively simple. It could be a different story for simpler engines.

[Rebel]

http://www.top-5000.nl/prodeo187x.htm

Cool stuff! I wonder if modifying the PST and material values is enough to be classified as a potential clone by the dendogram people. It would be interesting to test ProDeo with personality X for correlation to engine X, and see what comes out of it.

Interesting thought.

If the conclusion is positive, then it demonstrates that the statistical detection of clones is dubious: an honest programmer who writes the entire engine and borrows PST is considered a "cheat", while the real "cheat" will be clever enough to change the PST and come out clean

I remember Richard explaining the high similarity percentage of Critter with Ippo because the use of the Ippo piece values and PST's So now that the Ippo values can be used ProDeo becomes an Ippo clone? Now that would be real funny. I don't believe it for one moment.

Adam, where are thou ?

[Rebel]

Perhaps I will put my PST in an external binary file. That way, and if we can agree on the binary format for PST+Material, our engines can load any kind of "personality", and people can contribute personality files, etc. Also, perhaps bishop pair should be part of the material values.

Sure, if programmers are willing to agree on a standard. Count me in.

[mar]

Perhaps I will put my PST in an external binary file. That way, and if we can agree on the binary format for PST+Material, our engines can load any kind of "personality", and people can contribute personality files, etc. Also, perhaps bishop pair should be part of the material values.

Sure, if programmers are willing to agree on a standard. Count me in.

I too think this is a good idea. When people create personalities the first thing they do is change material values.
I'm not sure if anyone would bother to change psq tables though.
When I did personalities in cheng3, I exported almost all other parameters I had,
but this probably only confuses people. They don't want complicated weights with exotic names,
they want global weights like mobility etc.
However I would probably opt for a text format instead of binary (consider endianess, future extensions, readability by humans [and thus making it easy to make changes]).

[Rebel]

Here are the results of a (short) robin tournament with all the imported Piece values + PST's (into ProDeo) from the other thread.

Code: Select all

No. Engine            1     2     3     4     5     6     7     8     9  Score
------------------------------------------------------------------------------
 1 ProDeo1.87      xxxx  14.5  17.0  21.0  18.5  14.0  16.0  15.0  22.5  138.5 /  240 (57.71%)
 2 Hakkapeliitta   15.5  xxxx  13.0  13.5  15.5  20.0  17.5  20.0  15.5  130.5 /  240 (54.38%)
 3 DiscoCheck      13.0  17.0  xxxx  14.0  14.0  14.5  18.5  18.0  18.0  127.0 /  240 (52.92%)
 4 Rodent           9.0  16.5  16.0  xxxx  15.0  15.5  15.5  16.0  17.0  120.5 /  240 (50.21%)
 5 Ippo            11.5  14.5  16.0  15.0  xxxx  18.5  14.0  14.0  14.5  118.0 /  240 (49.17%)
 6 CPW             16.0  10.0  15.5  14.5  11.5  xxxx  16.5  16.0  17.5  117.5 /  240 (48.96%)
 7 Leonidas        14.0  12.5  11.5  14.5  16.0  13.5  xxxx  16.5  17.0  115.5 /  240 (48.13%)
 8 Adam_Hair       15.0  10.0  12.0  14.0  16.0  14.0  13.5  xxxx  16.0  110.5 /  240 (46.04%)
 9 Ice              7.5  14.5  12.0  13.0  15.5  12.5  13.0  14.0  xxxx  102.0 /  240 (42.50%)

No surprise to see ProDeo on top as its own values are in balance with EVAL, the others are not.

[Ajedrecista]

Hello Ed:

Here are the results of a (short) robin tournament with all the imported Piece values + PST's (into ProDeo) from the other thread.

Code: Select all

No. Engine            1     2     3     4     5     6     7     8     9  Score
------------------------------------------------------------------------------
 1 ProDeo1.87      xxxx  14.5  17.0  21.0  18.5  14.0  16.0  15.0  22.5  138.5 /  240 (57.71%)
 2 Hakkapeliitta   15.5  xxxx  13.0  13.5  15.5  20.0  17.5  20.0  15.5  130.5 /  240 (54.38%)
 3 DiscoCheck      13.0  17.0  xxxx  14.0  14.0  14.5  18.5  18.0  18.0  127.0 /  240 (52.92%)
 4 Rodent           9.0  16.5  16.0  xxxx  15.0  15.5  15.5  16.0  17.0  120.5 /  240 (50.21%)
 5 Ippo            11.5  14.5  16.0  15.0  xxxx  18.5  14.0  14.0  14.5  118.0 /  240 (49.17%)
 6 CPW             16.0  10.0  15.5  14.5  11.5  xxxx  16.5  16.0  17.5  117.5 /  240 (48.96%)
 7 Leonidas        14.0  12.5  11.5  14.5  16.0  13.5  xxxx  16.5  17.0  115.5 /  240 (48.13%)
 8 Adam_Hair       15.0  10.0  12.0  14.0  16.0  14.0  13.5  xxxx  16.0  110.5 /  240 (46.04%)
 9 Ice              7.5  14.5  12.0  13.0  15.5  12.5  13.0  14.0  xxxx  102.0 /  240 (42.50%)

No surprise to see ProDeo on top as its own values are in balance with EVAL, the others are not.

There is not a PGN file available, but just for give an idea to other readers, here is a list of rating performances of your RR (it should be very similar to EloSTAT performances):

Code: Select all

Round Robin with  9 engines and    240 games per engine.
Total number of games:      1080 games.
 
 Engines:     Performance:     Score:

Engine 01:        47.24       57.71 %
Engine 02:        26.66       54.38 %
Engine 03:        17.75       52.92 %
Engine 04:         1.26       50.21 %
Engine 05:        -5.07       49.17 %
Engine 06:        -6.34       48.96 %
Engine 07:       -11.41       48.13 %
Engine 08:       -24.12       46.04 %
Engine 09:       -45.95       42.50 %

Mean of ratings:     0.00 Elo.

I choosed an average rating of 0 randomly. The important thing is that the maximum difference is around 93 Elo (you should get different results from Bayeselo and from Ordo).

Regards from Spain.

Ajedrecista.