Author: Steven Edwards
Date: 15:44:39 03/15/04
Go up one level in this thread
On March 15, 2004 at 05:48:05, Sune Fischer wrote: >On March 15, 2004 at 05:06:09, Steven Edwards wrote: >Wouldn't it be possible to construct a simple testbed to quickly try out a few >of the ideas and see if they have merrit? No. There is too much foundation work needed to get even a simple plann exercise to work. >>My effort, Symbolic, has a search space. Its search space is a plan space and >>not a position/move tree. There is a position/move tree, but the plan space >>search wholly governs its growth. The tree is not for discovery, but for plan >>verification. Absolutely no nodes in the tree are expanded unless the plan >>search process directs such expansion, one node at a time, and only for good >>reason. The node count and the node frequency of the position/move tree is >>merely an epiphenomenon of the actions of the plan search and so cannot be >>compared to similar metrics of a traditional A/B searcher. In fact, given a >>goal of modeling human behavior, an excessive node count or node frequency is >a >>sure sign of failure. > >How does this work if winning a piece requires a 4 move combination? Inital pattern matching sugeests the possibility of a specific successful attack further pattern matching refines the idea and fills in details; a plan is constructed to persue the idea; the plan is verified via a limited tree search. >How do you make it search down the first few of those moves, moves which at >first glance might even seem silly? Because there is no "first glance". Symbolic is not a Shannon type B program. The moves are the last thing it looks at, not the first. >I can see how you could generate objects of attack, but how do you effectively >search down the "building up the attack" tree? Plans can be extended and refined while being verified. If a plan is making good progress, then the program can decide to spend extra resources on further verification including fault detection and repair. >Each move must be part of a greater plan, on its own it might seem like a silly >move. True. >How do you make this connection and utilize it? A good working plan has enough detail to get through the rough spots when something unexpected happens during the verification process. You might enjoy reading Wilkins' paper on Paradise; he has a good description and several examples.
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.