Author: Michael Yee
Date: 11:45:12 05/25/05
Go up one level in this thread
On May 25, 2005 at 02:12:13, Steven Edwards wrote: >On May 24, 2005 at 18:49:05, Michael Yee wrote: > >>Schaeffer essentially computes piece square table "bonuses" that depend on the >>root position (e.g., plans like "king-side attack" or "try to occupy f5 with >>your knight") and get added to the score of a branch in a path dependent way. >> >>Some questions: > >>(2) Does anyone know how Schaeffer's Planner compared to Wilkin's PARADISE (in >>terms of playing strength)? > >The only strength data point on Paradise is that it performed better on tactical >problems than either Chess 4.x or Tech II (given a significant time advantage >(45 minutes vs. 5 minutes per position). If the ChessLisp interpreter had been >available, the time advantage would have been unneeded. > >My understandng is that the output of Schaeffer's Planner was limited to a set >of twelve piece/square tables used in leaf node evaluations. This was generated >once, at the root node only. Paradise generated a relatively complex plan that >was then used as a search control program. Plans would also be generated at >other nodes when needed, and a variety of other somewhat advanced techniques >(e.g., a causality facility) were used to assist with the search. > >Schaeffer's root preprocessing work is interesting in that it's initial phase of >long term goal selection is similar to Symbolic's KsSurveyor theme selection. >Any planning program will need such a component. Symbolic has the capability >(and the requirement) to potentially apply the theme selection at any node, not >just at the root. > >Generating piece/square evaluation tables at the root and expecting them to be >of much use a dozen plies later is unrealistic. > >>(3) Is anyone experimenting with this or other types of path dependent eval? >> >>Steven is of course working on long-range planning. And I think Uri has some >>path dependent stuff. But have most people pretty much abandoned it? > >Instead of calling it "path dependent eval"', I prefer saying "root >preprocessing" meaning that it's a one time task done only at the root position. > Symbolic, like Paradise, has a recursive planner that can be potentially >applied at any node and is not limited to the root position. In fact, if an >obvious move is available in the root position then the planner will never be >invoked. Thanks for the explanations. I think Schaeffer's approach at least included the ability to detect when the root plan was no longer valid for a given branch (in which case the rest of that branch would be treated in the standard no-plan way). But the ability of symbolic and paradise to re-plan at any node sounds much more flexible and realistic. Michael
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.