Author: Tony Werten
Date: 23:14:12 01/19/06
Go up one level in this thread
On January 19, 2006 at 15:01:04, Joachim Rang wrote: >Hi, > >after a question below regarding the "superiority" of MTD(F) and the article on >Wikipedia about that I decided to modify the entry and added a sentence in which >I question the "superiority" of MTD(f) over PVS since it has "practical issues". >Naturally some guys over in Wikipedia now want to see proof of my statement and >want to know what kind of practical issues there really are. So I have two >questions: > >1.) Would you support the statement that MTD(F) is _not_ superior to PVS and >while save in some circumstances a few % of nodes has "practical issues" which >make it a less desirable choice? > >2.) Could you (briefly) name some of the "practical issues"? > >regards Joachim > >P.S.: If someone feels encourage to reformulate the article on MTD(F) and >Negascout (PVS) more knowledgeable please feel free to do so. ;-) I think MTD's main problem is to handle search instabilities. That wouldn't be a problem, but unfortunately, MTD is the most unstable search method. PVS is more stable, PVS without aspiration window is even more stable, plain alpha beta even more and minimax most (though unpractical). In addition, a lot of programming "tricks" (ie nullmove, pruning, reductions, lazy eval ) increase instability, making MTD look even worse. The reason why it looks ok in academic reasearch is IMO, that an academic chessprogram was used without a lot of these tricks. Take a chessprogram that only evaluates with piece square tables and counting wood, without nullmove, and no reductions/pruning, and nothing beats MTD. (I'm not sure about extensions, i think they tend to stabilize the search) Tony
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.