Author: Moritz Berger
Date: 02:25:32 08/14/98
Go up one level in this thread
On August 14, 1998 at 04:39:52, blass uri wrote: >This is a wrong claim against fritz5 from fritz5's user guide in page 46: > >"Please note that it does not make a lot of sense to make the program run for >many hours to analyse a position if the hash tables are full. The extra time >hardly compensates for the dramatic reduction in speed. It may be more >advantageous to use the "Correspondence analysis" function described in section >6.10 page 57" > >The reader may think that fritz5 does not earn from time like other programs >and it is better to buy another program to analyze a position for hours >when the truth is that fritz5 earns from time more than many programs. > >I found that the correspondence analysis is not a good idea because >it can miss surprising tactical moves fritz5 can find if it run for hours. Both methods have advantages: Due to search complexity, deeper search on one position will reveal less and less improvements about the quality of a preferred move. Fritz does heavy selective pruning and might miss best moves with increasing likelyhood the deeper in its search tree it encounters a critical position. An instructive example is to let Fritz analyze a position e.g. for 10 minutes. Then write down main line and evaluation and enter the proposed moves. Let Fritz analyze the position for 10 seconds. You will very likely get a more accurate score for the final position from the 10 second search than you got from the 10 minute analysis. The most sensible approach is trying to allocate sufficient time per move (e.g. 10 minutes on PII-400 with 128 MB hash tables would be my first guess) and use the correspondence analysis feature with a narrow branching factor (I would prefer 2 over 1 in the tree and maybe 3 in the initial position to include alternate moves with close scores to the one found within the 10 minute search). What's hash tables got to do with it? Well, considering the idea behind correspondence analysis (logarithmic expansion of the exponential search tree besides including or excluding moves specified by the user and exploring several alternatives), increasing the "sufficient search time per move" I mentioned above yields comparatively slower gains in depth (implicating accuracy) when hash tables are "exhausted". So hash table size is *one* constraint when determining the best search time per move as with smaller hash tables increasing the branching factor for analysis might yield potentially better results. I hope this explanation is clear, at least you can see from my humble attempt why the authors of Fritz' produced only such a vague hint that hash tables are involved when you're making the tradeoff branching factor vs. search time to analyze a position as well as possible within a given time (e.g. what's best when you plan to give Fritz' one night per position). >a dissapointed buyer of fritz5 because of the correspondence analysis >of fritz5(I suppose s(he) did not try to give fritz5 hours) can tell friends not >to buy fritz5 and chessbase will lose from it. > >I am not discussing about the level of fritz5. The problem affects all other programs (Fritz engines or not) in a similar way. >I simply do not understand why chessbase hides the fact Fritz5 earns from time >when the hash tables are full. Fritz gains of course from longer pondering times. Without the powerful correspondence analysis feature, there would be no alternative for overnight analysis than to simply have Fritz chew on one position as long as possible. But as I tried to outline above, moving the root position deeper down in the analysis tree is a nice trick to cope with the exploding search tree and choked hash tables. By including not only the (one) best move when "climbing" the tree, you ensure that better moves that are close in score in one position will also get a closer examination with increased search depth (+1 ply per step upwards in the tree). This implicates the main disadvantage: Spectacular sacrifices will most likely escape this analysis if they are not found within the search horizon of each individual move. The only workaround is to enter them as variants in the root position to force Fritz to analyze them too or to default to conventional "infinite" analysis. Hope this helps Moritz
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.