Author: Guido
Date: 07:39:43 04/29/03
Go up one level in this thread
On April 28, 2003 at 18:34:11, GuyHaworth wrote: > >Well, I haven't read the paper you cite and am not familiar with their >'language', but I will say something about retrograde EGT(B) generation. > > >There are two ways of generating EGTs, both perfectly ok, though differing in >efficiency. > >The 'Edwards/Nalimov' way looks at every (say, wtm) position - and for each >position examines its successors. It backs up the best option for White - which >might be 'nothing known' or 'win in N1'. This is not to say that, on the next >occasion, this 'win in N1' will not be improved to 'win in N2' with N2 < N1. > >The 'Thompson/Wirth/Wu-Beal' way is to literally 'retro' a 'frontier' of 'newly >evaluated positions'. So, if we back up btm losses, we know automatically that >they back up to wtm wins - and the only question is whether they are quicker >wins than already found. > I succeeded in solving EGTBs by both the methods. So I could compare not only the tables generated but also the efficiency of the methods and I found that the 'Thompson/Wirth/Wu-Beal' way (having read nothing about, I presume to have reinvented this method) is 3-4 times faster than 'Edwards/Nalimov' way. This factor is obtained for some 5 men endings, while I think it should be higher for 6 men and over. Moreover the advantage is bigger for the most interesting endings, in general the well balanced ones, where wins and losses are few and the calculation requests more cycles and time. This retro approach is nevertheless more complicate to implement, in particular when it is necessary to use disk memory instead of RAM. I know DTC and DTM, but what are DTZ and DTR? Ciao Guido >With this fully retro approach ... For DTC, and DTM if you use the Wu/Beal >technique of not backing up 'mate in N' until cycle N, and DTZ if you are >cleverer still ( but this hasn't been done yet), you can say that wins in N will >be set in cycle N so, in fact, they never get reduced. They just get their >value set once and it is correct. > >So, the integrity of EGT generation depends on how you manage successors-of-P >(or in the retro approach, predecessors of P and the known set of 'latest >results'). > >The fundamental reason these methods work is that they set interim depth values >which, if necessary, are open to being reduced. > >If you have the efficiency that you know the 'interim' depths cannot be reduced, >e.g. because DTC or DTM = cycle_number when set, then you do not need to try to >reduce them. > >I hope that is clear. > > >I have to say that the reasons why EGT generation works are non-trivial, and I >once tripped up and defined an incorrect algorithm for DTR calculations. So I >do sympathise with your question. > > >Guy
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.