Author: Stuart Cracraft
Date: 20:49:26 08/07/04
Go up one level in this thread
On August 07, 2004 at 23:38:19, Robin Smith wrote: >On August 07, 2004 at 23:27:59, Stuart Cracraft wrote: > >>On August 07, 2004 at 23:23:00, Robin Smith wrote: >> >>>Fortress draws have been the Achilles heel of programs for years, and probably >>>will be for years to come. This weakness will be very hard for programmers to >>>rectify. But it occurs to me that for analysis (not playing), this and other >>>program weaknesses could be partially corrected. If for specific positions human >>>evaluations could entered and stored in hash tables, with some kind of don't >>>erase this position or evaluation flag, the strengths of humans and computers >>>could be more readily combined. >>> >>>For example, in a post below, Uri analysed the following postion: >>> >>>[D]3r2rk/3n1pp1/2p1b2p/3q3P/pp1PNQ2/2P2P2/PP6/KB4RR w - - 0 29 >>> >>>Uri's analysis went 29.Rg6 fxg6 30.hxg6 Rge8 31.Rxh6+ Kg8 32.Rh7 Bf5 33.Qh4 Bxg6 >>>34.c4 Bxh7 35.cxd5 Bxe4 36.Bxe4 Nf6 37.Bg6 cxd5 38.Bxe8 Rxe8 >>> >>>[D]4r1k1/6p1/5n2/3p4/pp1P3Q/5P2/PP6/K7 w - - 0 39 >>> >>>Here, while computers see a large advantage, a human can spot the possibility of >>>a fortress, could verify the fortress does indeed hold, and then enter this >>>position's evaluation as 0.00. Then backing up from here, this info could be >>>retained in hash tables, and a program could quickly spot the improvement for >>>White: 36.fxe4 cxd5 37.e5 I don't think any commercial programs allow what I am >>>talking about, but perhaps one of the freeware engines does? And if not, how >>>difficult would it be to add such a feature, for analysis? It would be very >>>powerful. >>> >>>-Robin >> >>We understand your point but as always the problem is implementation. >>Placing the full position hash for these permanently in a file to be >>read in at startup, etc. with never-replace flags, is too specific >>a solution. >> >>You need something that generalizes a lot better than this. >> >>If you can reduce it to a large series of pawn structures that qualify, >>then you can have guidance for the program that might make sense. >> >>Even then I am doubtful of it. This has been done and it isn't too popular. >>If so, why not? Is the rate of return too poor or simply too esoteric, >>theoretical and not practical? >> >>The "mouse" program of Slate which saved poor search results into a >>permanent hash table entry, comes to mind, but is far too specific. >>Kalme and Zobrist's advice taker too primitive and rather weak requiring >>intensive Master for questionable results. There are others. >> >>I'm sure you'll get a long list now. I would like to see something like >>this succeed though. >> >>Stuart > >Hi Stuart, > >I am not talking about permanently storing this information, only >semi-permanently. There are too many positions to store them all permanently, >and trying to do so would serve little purpose, since the identical position >would rarely reoccur. Thus my idea wouldn't help a program's playing. Nor was >that my goal. But when someone is trying to analyse a specific position, I think >something very effective could be implimented. It just requires a programmer to >care as much about chess engines as analysis tools as they care about them for >playing strength and winning tournaments. Unfortuanately for someone such as >myself, who is more interested in analysis, winning tournaments seems to be the >most popular programmer goal. > >-Robin I liked Zobrist and Kalme's research many years ago into this subject. Kalme would provide a series of patterns to guide the program. Not specific positions. Another example is David Wilkins program for tactical problems. Good to hear this kind of interest is still alive in this search-exhausted brute-force selective-days. I am pro-patterns and pro-learning! But the implementation is bedeviling. Stuart
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.