Deep Fritz 11 eval

[hyatt]

I'm hoping that his analysis is only dealing with a single "false-positive" here. Because it has provided two kinds of information. First, it has (if it can be trusted) shown that several commercial programs are completely original. As opposed to a few on Talkchess or the Rybka forum that claim all commercials are code-copying cheaters. I find it hard to imagine that one could copy any significant amount of code, and still look "unrelated". But we will see. However, secondly, it has suggested that Fritz 11 is related to strelka/et. al. The interesting thing is that it suggests that fritz 11 is not related to fritz 10, and that represents a huge departure from the rest of his results. If you look at "families" in his graph, different versions of the same program always show to be very close relatives. Which is, to me, the expected result. That Fritz 11 is so related to strelka is interesting. That Fritz 11 is so unrelated to fritz 10 is really unexpected. However, commercial authors are always quite secretive, so I do not know if there are any known specifics about the differences between F10 and F11. For example, adding reductions to a program that didn't have them might suddenly make that version appear much closer to others that use reductions as opposed to the original that did not. If this question can be answered, it is possible that this could be a first-approximation to decide which programs are suspicious and need further examination. I'm still convinced that the final determination has to be source code (or RE'ed source if necessary) comparison to make sure we are not looking at the same idea with two different implementations, which would be perfectly acceptable...

But it would certainly be nice to have some sort of filtration process to eliminate RE analysis of so many engines, which is a pain... without overlooking those that are potential problems...

[thorstenczub]

i doubt fritz programming is done anymore by frans morsch himself.
IMO with fritz has happened similar things that happened to many commercial chess programs, one day the programmer in charge has sold his sources to the company who wants to exploit the sources. so the companies can develop the program on their own.

i doubt frans programs fritz anymore. IMO the programmers in the chessbase company develop the program further.

so by doing this, they maybe put stuff from the cloned programs into fritz, to get it within the top 10 or not lose the contact to the leading programs on top.

whatever.

i remember dave kittinger once said that he was not in charge for the latest NOVAG programs anymore. i guess novag bought the code and asked somebody else to develop
a new machine by modifying the sources.

[BB+]

I might note that the ponderhit numbers from CCRL are at 40/40m, while DD specifically chose short searches (0.1s) to emphasise evaluation. I'm not sure what to make of them. The CCRL numbers also include "eval difference", and F11/R232a is remarkably close -- then again R2.x dominates the "top 30" there, maybe because they were tested so much (you buy more lottery tickets, you have a better chance of winning...). The R12/R1 eval difference is the greatest "in-family" numbers, but inspection of the two shows them to have almost the identical evaluation function (numerology included). One of the first things an experimentalist should do is determine what exactly their tools are measuring. In the case at hand (with both CCRL numbers and DD/dendrograms), I can't say that have much understanding of them yet.