Author: Robert Hyatt
Date: 07:16:39 05/09/01
Go up one level in this thread
On May 09, 2001 at 07:24:35, Graham Laight wrote: >Just a thought - I don't know whether it's going to be useful or not. > >Suppose you switch off the opening books, and play the opening to the middlegame >from the engine. > >What I would like to know is whether there is a correlation between the depth of >search and the proportion of node positions which have already been seen. > >Would the hash hit statistics be able to show this? They could show it but they would be wrong. We currently has raw positions, and we _know_ we are wrong in doing so since we don't hash the move path to reach the position, so that we overlook draws by repetition and 50-moves that we just accept as an error term. If you factor in move path information, I think the transposition hit rate would be pretty constant no matter how deep you search, compared to what we see today due to faulty hashing. > >If this can be discovered, I suspect that we're going to find that the deeper we >search, the higher the proportion of nodes we generate will have already been >seen. > >In terms of estimating the size of chess, what would be REALLY interesting would >be to see the shape of the following graph: proportion of repeated positions >plotted against depth of search. See above. With today's programs, this might look good. But it would be both wrong and misleading and if we _could_ (say) search to a depth of 300-400 plies, we would _all_ be hashing correctly or we would be drawing won games right and left. > >If the graph turns out to be a straight line (a normal correlation), then chess >is smaller than we all think. If the graph turns out to be logarithmic, then >chess is very big, and it's going to be difficult to solve. It is _very_ big. > >Does anyone out there know what the shape of the graph of hash hits against >search depth is? > >-g Depends on whether you mean "todays incorrect hashing scheme" or "a real hashing scheme".
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.