Henk wrote:If you are using chess engines only for analysis then you are not interested which engines find best moves within few minutes but you want to know what is best move even if that takes many hours.
So results of tournaments with ultra long time control may be more interesting for serious chess players.
mostly, the engines make the same mistakes at LTC they make at 1 min.
current SF is better at 1 min. than SF 7 at 1 hr.
make the conclusions yourself.
Actually I was thinking about 12 or 24 hours per move.
TC does not matter.
give SF 2 days, 10 days or 100 days, it will still prune the same lines and hit the very same eval nodes.
That means SF is not suitable for it prunes too much. For instance if you would use plain alpha beta you can assume moves getting better when giving more time for it searches deeper and no moves are pruned.
plain or not, you need good eval.
similarly, Kasparov spots in a second what other players would need minutes or even hours to spot.
software is always more important than hardware/available time.
Henk wrote:If you are using chess engines only for analysis then you are not interested which engines find best moves within few minutes but you want to know what is best move even if that takes many hours.
So results of tournaments with ultra long time control may be more interesting for serious chess players.
I agree with you. I would like to see the chess at correspondence time controls using a raft of 64 core monsters with access to enormous RAM and tablebase resource.
If you had enough 64 core machines, you could run all the pairings simultaneously.
There are, unfortunately, two problems. The first one is the millions of dollars it would cost. The second is that nobody has any patience. Even if you had 128 machines and 32 entrants, the contest could still take over one year to complete.
So, tragically, it will never happen.
does not make much sense to me.
Rybka on 256 cores currently is definitely weaker than current SF on a single core.
and those 256 cores costed a lot of money, while SF is free.
so, I would concentrate on improving software, rather than using tremendous hardware/VLTC.
Henk wrote:If you are using chess engines only for analysis then you are not interested which engines find best moves within few minutes but you want to know what is best move even if that takes many hours.
So results of tournaments with ultra long time control may be more interesting for serious chess players.
I agree with you. I would like to see the chess at correspondence time controls using a raft of 64 core monsters with access to enormous RAM and tablebase resource.
If you had enough 64 core machines, you could run all the pairings simultaneously.
There are, unfortunately, two problems. The first one is the millions of dollars it would cost. The second is that nobody has any patience. Even if you had 128 machines and 32 entrants, the contest could still take over one year to complete.
So, tragically, it will never happen.
There is a speed up. If estimated Elo difference between two engines is more than 100 points one can give the engine with highest rating less time. Assuming less time does not make it play better.
From what I gather from correspondence chess practice, what usually happens is you look at a few moves that seem sensible and explore those branches to some depth (say ~40) and see what comes up.
It doesn't take forever and it's better than just exploring the principal variation to depth 50 or so.
So VLTC games probably aren't the best way out. Not for CC at least.