Author: Steven Edwards
Date: 16:24:59 04/04/05
Go up one level in this thread
On April 04, 2005 at 18:16:10, Roger D Davis wrote: >On April 04, 2005 at 03:07:13, Steven Edwards wrote: >>Symbolic: Status report 2005.04.04 ... >Looks like your project is really coming along....big congratulations! Was >wondering if you'd speculate on a few things: 1) Once Symbolic is up and >running, how might it be possible to "dumb it down" for players of various >levels of ability? 2) Also, will it be possible to give the program various >styles of play? 1. Well, first it has to be smartened up prior to any dumbing down, and that's a long ways away. I suppose one idea is to artificially limit its planning efforts, its plan exproation efforts, or both together. 2. Symbolic's style, assuming it has one, is based on its plan generation. In a fashion, it searches through a plausible plan space sowewhat like a Shannon Type B program searches through a plausible move space. A chess position space is searched only after a plan is manufactured. So the playing style will be dependent on what data is used to suggest high level plans and how multiple alternative subplans are evaluated for ranking prior to position tree search. >These are both issues faced by programmers in the Alpha-Beta paradigm, but since >you're using another approach, you might have different solutions, and your >solutions might be better or more elegant. For example, creating styles now is a >matter of changing the weights of various scoring terms to see what happens. How >would creating styles proceed with Symbolic? Since traditional A/B searchers have nothing to do with planning, there's unlikely to be any ways in which they can be incrementally improved by adopting techniques explored by Symbolic. In Symbolic, there's not a vector of simple scalars anywhere that need only be tuned in order to change program behavior. Instead there is (or, will someday be) a large library of Lisp coded productions (i.e, symbolic re-writing rules) and corresponding sequencers that together embody the chess knowledge of the program. This is the major similarity between my program and Wilkins' program Paradise. Any emergent playing style is not programmed in from the start; instead, it is an epiphenomenon of the plan formation process. And what of the plan formation process? Surprise! It has nothing to do with chess! It is a domain independent pattern recognition scheme that tries to produce a model of its input (a chess position in this case) that expains the relationships of the parts to the whole. Of all other common algorithms, the idea is closest to a natural language chart parser (c.f. Winograd's _Natural language Understanding_). With a chart parser. the results are the various possible syntax parse trees for possibly ambiguous input. For Symbolic, the results are the various plans that may be useful for controlling a position tree search.
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.