Author: Robert Hyatt
Date: 11:43:48 02/16/04
Go up one level in this thread
On February 16, 2004 at 10:31:03, Steven Edwards wrote: >Symbolic: The TNS (Thousand Node Search) > >The idea of limiting the cognitive search in Symbolic to under a thousand nodes >is based upon psychological studies that suggest top level human chessplayers >usually visualize between 100 and 1,000 positions per move in complex >middlegames. My personal time control upper limit preference for non-blitz >chess is a minute per move, and so the resulting target figure for node >frequency is about 20 Hz. I think you are starting off here using an unsound assumption. "100 to 1000 positions per move" is probably nowhere near right. There is a difference between a human mentally moving pieces around, and his comparing them to pattern-recognition information that in itself is the result of searching significant amounts of tree space. Who knows _what_ I actually do after thinking a few minutes and moving the pieces around in my head, to decide 'this position is one I want to reach." Did my "static evaluation" fold in a bunch of past experiences via pattern matching? IMHO picking some number like 1K is just picking a number like 1K, not that 1K is more or less meaningful than 100 or 10K... trying to quantify how many "positions" a human searches is pointless until we know how a human really "searches". To date, we have no idea. this probably won't change for many years, until all the marvelous abilities of the human brain have been analyzed and understood. > >One idea here is that the target frequency remain somewhat invariant of the host >hardware. On faster machines, the effort expended on non-search chess knowledge >can be increased, while on slower hardware, it can be lessened. Similar >throttling can be applied for different time controls on the same hardware. > >A distinction here between the large pool of iterative A/B searchers (and their >hardware brethren) vs programs like Paradise and Symbolic is the purpose of the >search itself. For the descendants of Slate and Atkin program Chess 4.x, the >main purpose is discovery. For Paradise, and for Symbolic to a slightly lesser >extent, the main purpose of the search is plan verification. > >It is important to note that Symbolic is not a "selective search" program is the >commonly used sense of the phrase. A selective search program is one that >employs the Shannon type B strategy of reducing the full width search at each >node by applying a plausibility filter or by having a plausible move generator. >Shannon type B is the same as type A in that the purpose of the search is >discovery; the topography of the resulting depth first search trees may differ >in mean height and width, but the reasons for searching any particular node are >the same. > >Symbolic, like Paradise, should only expand a search node (i. e, move generation >plus selection) if it has a good reason to do so. The phrase "good reason" is >somewhat vague (at this point in development), but the one thing it does not >mean is "there's still time on the clock, so let's try another move/another >iteration". Instead, Symbolic will always have an active plan that will >determine which nodes to expand. In some cases, multiple moves at a given node >will be conforming to the current plan. In others, perhaps no move will >suffice, so the plan must be modified or abandoned. As Symbolic keeps the >entire search tree available at all times, a node may be revisited with >different plans. > >And although the exact implementation details of the above phrase "good reason" >are not well defined at this point, it is guaranteed that each "good reason" >will have a natural language representation. Thus, another reason for the TNS >node count limitation: Symbolic will produce an explanation audit trail with all >of the good reasons, in English, for each decision made during the search and >this document has to be easily readable (by me) for the purposes of tutoring the >program. A multi-megabyte dump will not be useful, but a five or six page >synopsis should work well.
This page took 0.01 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.