Author: Robert Hyatt
Date: 20:14:57 06/05/01
Go up one level in this thread
On June 05, 2001 at 20:33:58, Vincent Diepeveen wrote: >On June 04, 2001 at 22:45:44, Robert Hyatt wrote: > >>On June 04, 2001 at 18:25:31, Vincent Diepeveen wrote: >> >>>On June 04, 2001 at 10:59:41, Robert Hyatt wrote: >>> >>>>On June 04, 2001 at 06:44:26, Dan Andersson wrote: >>>> >>>>>Impressive branching factor. Do you use ETC to reduce the tree/graph? And do you >>>>>try to search smaller subtrees by finding irregular branching factors? I think >>>>>it would be interesting to see its search performance without Q-search. >>>>> >>>>>Regards Dan Andersson >>>> >>>> >>>>MTD(f) is a killer with material-only. You don't do _any_ re-searches at >>>>all, which means it searches a perfectly ordered tree. >>> >>>Noop it doesn't necessarily search a perfectly ordered tree depending >>>upon move ordering and nullmove implementation. >>> >>>Material can fall of the board by stupid moves for example. >> >> >>Try it before saying that. It is so close to perfectly ordered that it can >>be called "perfectly ordered" with no danger of being wrong enough that it >>can be measured. He said he had move ordering operational. Which means >>good captures first. That is enough for a material-only search. Throw in >>killers and hashing and it is so close to perfect it counts... > >No it is not. The reason why it is so good ordered here is >because Rudolf generates the simplistic pawn moves as first in >this position. > >those directly allow a nullmove then. > >If i generate however first Nf3 then Ne5 then Nd7 then my piece falls >of the board, then i will try for another 20 ply below that all kind >of stupid moves before i conclude that Nd7 is a stupid move. No _reasonable_ chess program is going to generate Nd7 _first_. That is the point. With mtd(f) all that is needed is that the _first_ move searched be accurate down the left-hand side of the tree. The hash table will take care of that until the _last_ ply of each new iteration. If it is wrong there, you look at 30+ extra nodes in one or two places. That is pretty accurate. > >tactical move ordering and is of *major* concern here. the many >available stupid pawn moves here are both for white and black the >reason that one can get so deep quickly here. The opening position is the reason he can get so deep. He _never_ has to do another root-position research. > >When talking about the number of legal moves there is not a major >difference here in the lines seen after a 30 ply search and >the lines seen by a search somewhere in the middlegame. However >a simple pawn move there is usually dropping material somehow >by some deeper tactics. A very good move ordering there would >give a way bigger depth as in the openings position because >in the middlegame already a few pawns are probably exchanged and >some pawns are placed against each other (so no legal moves allowed >by those pawns). After a few piece exchanges in the search the >number of legal moves again drops bigtime and hashtable works better. > >So theoretically in far middlegame one can search way way deeper >with material only search as in openingsposition. Totally wrong. From the opening position there are few tactics. In the middlegame, tactics is what the game is all about. In mtd(f) if you guess right on the material score, you win big. In middlegame positions there will be more re-searches and window errors to deal with. It won't be as efficient as from a known starting 0.00 position. > >Yet the fact that DIEP picks the light pieces there first and >pawns as second means that DIEP's move ordering sucks there, whereas >SOS rocks the boat compared to DIEP. > >The 0 bound score for alpha i get quick too. You don't get it as quick as mtd(f) does. you will spend over 1/2 of the total search effort just figuring out that the score at the root is 0.00, then you dismiss the rest of the moves more quickly. mtd(f) will find that 0.00 a _lot_ faster.
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.