Rybka 2.32 on an Athlon 1.2 GHZ could not be that high......

[pichy]

There's no way that this could be right, especially considering that Rybka 2.3.1 on a single Athlon 1.2 GHz is well over 2900 on the SSDF list, which is calibrated to the FIDE list. It is believed by some (mainly to be safe and conservative in estimating ratings) that the computer rating lists are inflated by a certain factor for ratings different from a certain rating. Based on various human vs. computer matches with various programs, it is reasonable to believe that the rating for which the SSDF list is correct for that particular hardware is somewhere between 2700 and 2800, and that ratings lower than this are too low, and ratings that are higher than this are too high. Anyway, being very, very conservative and assuming that the ratings higher than 2700 on the SSDF list are off by a factor of 2 from 2700, this would still put Rybka 2.3.1 at a rating of 2820 on a single Athlon 1.2 GHz, and on 2.2 GHz, this could become around 2860 (again, a conservative estimated increase), and on a dual, it would be 2890 (another conservative estimated increase). In reality, not being so conservative, the actual estimated rating that would equate to a FIDE elo would probably be at least 2950 or so.

[tjfroh]

Jorge,

I use Deep Rybka 3 on an Intel Core duo 2 E6850 using both processors. It is about two thirds as fast as Rybka 2.3.1 on a newer Core i7 975 using one processor core. Of course if you used Deep Rybka 3 on the newer machine then there would be a significant improvement.

Unless you do deep positional analysis you will still lose to the old Rybka 2 on an Athlon 1.2 GHz single processor if you don't play anticomputer chess.

Tio Timmy

[Sean Evans]

Probably around 2500 ELO on that box.

[Steve B]

Ay particular reason this thread is in the CTF?
If not i will ask one of the CTF mods to move it to the CCC(without leaving a shadow)
Regards
Steve

[pichy]

Ay particular reason this thread is in the CTF?
If not i will ask one of the CTF mods to move it to the CCC(without leaving a shadow)
Regards
Steve

I don,t know why it is on two places, you should remove the other post which is similar to this one without leaving any trace of it.

[Mike S.]

(...) which is calibrated to the FIDE list.

I think that a long time ago, there were man vs. machine games being played in Sweden for such a calibration, but with players around 2000 or maybe 2200 Elo, and less. I think that calibration is meaningless since a long time, with comps and engines becoming top GM strength and beyond.

Maybe someone from the SSDF list is reading this and can give details on that calibration, and what remained from it.

(At least) once, the SSDF level was reduced by a linear factor, last time by 100 points, I think because the top ratings seemed to be unrealistic relative to FIDE Elo. I think these calibration attempts are quite hopeless anyway. I'd rather standardize the ratings by assigning 0 (or 10,000 or anything which signalises "non-FIDE rating" directly) to a mid-range reference engine.

For example, if typical ratings lists would range from -700 to +700, it is clear that there is no calibration to FIDE and all these misintrepretations as if that were FIDE Elos, could be avoided. I think CCRL has had such a concept in the beginning, and I am actually sorry that they discarded it.

Also, if the big important ratings list groups could agree about the reference engine, it would help to compare results. Currently they simply have different number levels, IOW rating differences are obviously not (or only to a small share) based on the different conditions. E.g. CCRL is generally ~50 points (as a rough guess for the average diff.) higher than CEGT. I think it's just due to a base level which is required as an input for the rating calculations, and that's standardized neither as it seems. That are all reasons for misunderstandings if people are not insiders of all that stuff.

For example, journalists who aren't chess and ratings insiders, would look at an engine list and at the FIDE list, and quote these numbers in comparison as if they would come from the same test method or -scale and same players pool. So, the very similar numbers level is in fact confusing people. It suggests a connection which is not there.

[Spock]

For example, if typical ratings lists would range from -700 to +700, it is clear that there is no calibration to FIDE and all these misintrepretations as if that were FIDE Elos, could be avoided. I think CCRL has had such a concept in the beginning, and I am actually sorry that they discarded it.

We did, but it was too radical for most people and was not liked. People liked to see numbers they were sort of familiar with

Also, if the big important ratings list groups could agree about the reference engine, it would help to compare results. Currently they simply have different number levels, IOW rating differences are obviously not (or only to a small share) based on the different conditions. E.g. CCRL is generally ~50 points (as a rough guess for the average diff.) higher than CEGT.

Our list is based on SSDF values back in Nov 2006. We took a list of engines common to both lists at that time, and set our base level / scaling so that that basket of engines had the same average rating on both lists. A 40/40 time control on AMD64 4600+ ponder was not a million miles away from 40/120 on an Athlon 1200, but obvious differences with books etc. Anyway, so our base level or scaling was based on SSDF values.