First decision: 8 cores each with ponder or 16 cores each without ponder. I favor 16 cores without ponder, partly because Clemens is running the ponder on test, no need to duplicate that. Also, in my opinion using twice as many cores raises the quality more than having ponder on, and also ponder on somewhat increases the luck factor as guessing the opponent's move is mostly luck.ouachita wrote:Larry - as you know, I'm planning a "LTC" 16-core 2687W, no ponder, event between H4ProB, K1142 and SF DD (or TBD), and ask for your, Marco and Robert's suggestions on format, along with anyone else's.
Second decision: Time limit. It's hard to improve on the TCEC time limit of 2 hours plus 30" which corresponds to a major time limit in human play and has about the ideal ratio of base time to increment. But then you are pretty much just rerunning TCEC for the top three finishers. If you want to have the highest quality games on record, go to 2 1/2 hours plus 30", which is fairly comparable witih the traditional forty moves in two hours on average.
Third decision: hash table size. Half of your total memory is reasonable for long games like this. using many cores.
Fourth decision: Tablebases. Well, since Komodo doesn't yet support them I would of course prefer that they not be used, but I'm not neutral. I understand that the Stocfish team volunarily voted not to use them in the TCEC final; I don't know their reasons. They probably improve the quality of endgame play a bit perhaps depending on which TBs are used and how fast is the access. In any case I think it's fine to have the GUI use TBs to adjudicate (or rapidly play out) the games.
Fifth decision: books or opening test suites. I would favor test suites preferably ending after around six moves or so to give the engines the most chances to outplay their opponents.
Sixth Decision: Color reversal for each opening. Do it!
Seventh Decision: Settings. Follow the recommendations of the respective teams as stated on the internet.
Probably I've forgotten something!