Author: Dann Corbit
Date: 11:02:13 07/14/00
Go up one level in this thread
On July 14, 2000 at 13:50:39, Robert Hyatt wrote: [snip] >Even if you build the book from only CAP positions, at each position where >you have the "best move's CAP score" what about all the other moves? If you >have 30 moves, you have a pretty good chance that the CAP move may well be the >best (and right move). Say 50%. But after 2 plies, you are down to 25% being >sure. If you compute .5^30 or so, you see that you are almost guaranteed to >choose a CAP line that has a bust somewhere deeper, without the score showing >you that the bust is going to happen... One way to deal with this situation is to use only "lead-pipe cinches" in forming extrapolated pv, pm, ce. For instance, suppose for some position q, the analysis is 15 plies deep. The analysis for record q insists upon choice 5 of the 38 possible choices from that position. One ply forward from here, we have 8 other records analyzed, one of which is the suggested move. If the suggested move has the same depth (15 plies) then whatever it says is the ce, pv, and pm is surely better. We can therefore update record q with that data. Examination of the other 5 records may reveal a better choice at 15 plies depth. If any of the records are only 14 ply or less, we just ignore them. We can iterate this process and also address holes (make a "most wanted" list of missing positions that ought to be analyzed, for instance). Eventually, through this method, we could obviously get much better data. I see no reason why an average ply depth of 19 plies should not be reachable. I would not replace the actual pv, pm, and ce. Rather, I would create new fields called extrapolated pv (epv), extrapolated pm (epm) and extrapolated ce (ece) and tag the record with those values. Hence, we have the very solid computer analysis and also speculative choices. A computer simulation could be run to see what mix of probability should be used in choosing one type over the other. It may even turn out that the extrapolated data is always better.
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.