Don wrote:
Are these tables used in the tree search, or just to find the best move from the root position?
Or does it depend on the program?
If the Tables are used at root or in the search is completly up to the program! Right now I dont know of any prog. which uses the Tables only at root by default - except Critter (but I do not pay much attention as I strongly believe that it doesent matter much for gameplay and elo performance!). Of course some use it earlier, some later ...
But: I do not use Tablebases in the GUI. The Enigne has to support it or it has to run without TBs.
Nonetheless, I already started the run with the GTB (on one comp) for Houdini, but on request I use the standard version and this will be the Version I include in the IPON.
Most likely I am courious enough to repeat it a third time with GTB AND the PRO version to see if there is a difference.
Everything on start in a few minutes!
Bye
Ingo
Is there any measurable difference in speed (NPS or time to complete N ply) between the two versions on your hardware?
This morning I discovered that I forgot to include the Gaviota Tablebases. Even if I only use the 4pc version the current run is not according to my normal standard. As it is interesting to see if there is any difference I will finish this run and start it imediately again WITH 4pc GTB. The latter will be included in my list.
As a side effect we have a nice test about the influence of Tbs!
Sorry for the inconvinience
Ingo
In 2,100 games there will be no side effect. In all my experience, usually TBases give from +3 to -3 Elo points, which are measurable in 30,000+ games with 2 standard deviations (95% confidence, I mean the sign).
I figure maybe you have an answer to this question, pertaining to Elo calculation:
If you look at Ingo's result http://forum.computerschach.de/cgi-bin/ ... 1#pid41321
can you explain why Houdini's calculated Elo (3016, resulting from Elostat or Bayeselo) differs so much from the average of the individual matches Elo (the so called Perfs, at right), which I calculated to be 3045 ?
Note that I expect the performance average to be too high because performance is not linear.
An extreme example is that 100-0 and 50-50 is going to give an infinite performance for the 100-0 part so it is certainly higher average performance than
75-25 and 75-25
I still suspect that something is wrong with the result because the difference from 3016 to 3045 seems to me too high and houdini got higher performance than 3016 in 18 out of 21 matches based on the data.
The versions produce the same PV, the node count is slightly faster on the STD version (2.5%), but for whatever reason the last depth on the Pro version is doubled, while the standard goes a little deeper.
Overall I would say there are slight differences. Whats the impact of this is unknown.
The calculation seems to be correct. The individual perfomances are calculated vs the former rating. E.G. 2800 of Shredder. while the overall performance is calculated vs the Elo average of all 21 opponents. This average is 2769 (2768.857 - in the given example)
80.67% vs a 2768Elo opponent is 3016 Elo.
What I want to say is, that the average of the individual perfomance is something completly different than the performance vs the average Elo.
I know that your mathematical skills are far beyond my capabilities but I knoww as well that the calculation done by the Shredder-Classic-GUI is allways right for Elostat and differs just slightly for BayesElo. In the current discussion it is even right for both. When I take the games and throw them into Bayes or Elostat I get 3016.
If this is "right" or not is an intersting discussion, but it is the way Elos are calculated. Nonetheless I still hope for something better which will be accepted ... !
