Author: Robert Hyatt
Date: 16:43:59 03/03/06
Go up one level in this thread
On March 03, 2006 at 17:01:40, Tony Werten wrote: >On March 03, 2006 at 14:06:12, Robert Hyatt wrote: > >>On March 03, 2006 at 02:42:48, Tony Werten wrote: >> >>>On March 03, 2006 at 00:36:19, Robert Hyatt wrote: >>> >>>>On March 02, 2006 at 02:14:56, Tony Werten wrote: >>>> >>>>>On March 01, 2006 at 16:05:06, Robert Hyatt wrote: >>>>> >>>>>>I am still using my old 12-bit history index <to><from> to index into these new >>>>>>values. I also still maintain one for white and one for black. I once tried a >>>>>>17 bit index <piece><to><from> but it offered no improvement to the basic >>>>>>history heuristic, but I have not tested that with this reduction stuff. >>>>>> >>>>>>Any ideas what others are doing here? I can think of several possibilities: >>>>>> >>>>>><piece><to> >>>>>> >>>>>><piece><from> >>>>>> >>>>>><to><from> (I am doing this now) >>>>>> >>>>>><piece><to><from> >>>>>> >>>>>>to name at least 4. First two seem too simplistic. Last one turns the history >>>>>>tables into pretty good sized arrays (2^17 for white, ditto for black). >>>>>> >>>>> >>>>>It does only once if you use this to index a second table which returns an index >>>>>only for actual legal moves. >>>>> >>>> >>>>Not sure I follow??? "actual legal moves"??? >>> >>>If you use piece-from-to (12*64*64=49,152 entries) a lot of your array can only >>>be filled with moves like Qa1-b8 and the like. >> >>OK. Now I understand what you were talking about... wasted space... > >Sorry, it made perfect sence to me :) > >Explanations tend to make more sence when you know the thoughts of the writer. >Unfortunately, this works best in the limiting case of reader==writer > >Tony > always. :) But in any case, I ran a bunch of tests this afternoon comparing <piece><to> against <to><from> and they were very close. But most of the time <to><from> was always slightly better, so I didn't make any permanent change... >> >> >>> >>>Using a secondary table wich only has an index if the move is legal, will result >>>in a moveindex range of about 4000. >>> >>>Now every table you use based on the moveindex only has to be 4000 in size >>>rather than 49152 when you base it on the move. >>> >>>It can be used for the obvious stuff (history table,fh stats etc) but also for >>>more interesting things (fast semi possible move detection ie does it have an >>>index, piecesquare values of a move etc) >>> >>>Tony >>> >>>> >>>> >>>>>Tony >>>>> >>>>>>Note that I am essentially factoring in wtm already since I have two separate >>>>>>sets of history tables for black and white.
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.