Author: Omid David
Date: 02:29:37 09/03/02
Go up one level in this thread
On September 02, 2002 at 20:37:35, Vincent Diepeveen wrote: >On September 02, 2002 at 13:42:18, Omid David wrote: > >>On September 02, 2002 at 13:34:16, Robert Hyatt wrote: >> >>>On September 01, 2002 at 17:08:58, Omid David wrote: >>> >>>>On September 01, 2002 at 12:08:21, Uri Blass wrote: >>>> >>>>>On September 01, 2002 at 11:55:32, Vincent Diepeveen wrote: >>>>> >>>>>>On September 01, 2002 at 10:20:08, Uri Blass wrote: >>>>>> >>>>>>if you search for aske plaat you will find his stuff on >>>>>>mtd online probably. i'm amazed that you don't understand >>>>>>that Frans is using nullmove. >>>>> >>>>>I know that he is using null move but I do not use null move when I search the >>>>>line that is in the previous pv because I consider it a waste of time. >>>>> >>>>>Null move is for prunning illogical lines. >>>>> >>>>>The pv cannot be illogical line so the only case when I can save nodes by null >>>>>move pruning when I am in a pv line is when the position is zugzwang. >>>>>In other words I can save nodes only if null move pruning is wrong. >>>>> >>>>>I understood that MTD says that the pv may be wrong but the first ply of the pv >>>>>is always right so it does not make sense to prune after Nc6. >>>>> >>>>>Uri >>>> >>>>It depends on what you want; if you want "the first move, only the first move, >>>>and nothing but the first move", then use MTD(f). But if you also want the PV >>>>(as most of us do), avoid it. >>>> >>>>The use of null-move pruning follows the logic behind each method: In MTD(f), >>>>you only want the first move, so you are free to use null-move pruning even in >>>>the PV, something you shouldn't do in regular NegaScout/PVS. >>> >>>You are not thinking this thru clearly. null-move on the PV makes perfect >>>sends. You are searching one ply different. It might suddenly change things >>>in a significant way and null-move is the fastest way to dismiss a bad line, >>>whether it was part of the PV from the previous iteration or part of some >>>non-root move that is hopeless. >>> >>>You should try it in a program, doing it everywhere and only doing it on non- >>>PV searches. I think you will be surprised. I was. The comments in main.c >>>should explain when I made this change myself after someone (Bruce I think) >>>suggested it and testing proved him correct. >>> >>> >>> >>>> >>>>P.S. >>>>Aske Plaat first introduced MTD(f) in his 1995 article "An Algorithm Faster than >>>>NegaScout and SSS* in Practice" available at >>>>http://www.cs.vu.nl/~aske/Papers/hk.pdf >>>> >>>>For a more intuitive explanation, look up http://www.cs.vu.nl/~aske/mtdf.html >>>> >>>>Plaat suggests that MTD(f) is faster than NegaScout, but he researched only >>>>fixed-depth full-width trees. I haven't seen any publication concerning MTD(f)'s >>>>behavior in variable depth trees (e.g. using null-move pruning). >>> >>> >>>I suspect it is "break even" once you get it right. NO way to avoid at least >>>two searches, and, in general, more than that. Which means that with a program >>>with a complex evaluation, the researches are going to cause problems and make >>>it catch up to PVS or even pass it. >>> >>>I think that for normal programs, they should be equivalent if they are both >>>implemented with the same skill and development time. >> >>Actually I think MTD(f) needs more time for fine tuning. For example, you have >>to adjust the evaluation function to a good degree, since the behavior of MTD(f) >>can change to a great extent depending on the eval function; while PVS isn't >>that dependant on the eval function. > >It is the opposite. MTD works for idiotic random tree searches (especially >not having a qsearch like in aske's experiments that's the case). This confused me for some time: didn't Aske use at least qsearch? He speaks of fixed-depth full-width trees, but even there (in such a brute-force framework) lack of qsearch doesn't make sense. What's the point if I capture your piece and don't see the recapture in the next ply... >Also >for programs who are very bad in ordering moves and have huge overhead >otherwise to determine a good PV (which i get with PVS directly out of >hashtable already, like if score drops it doesn't take that much >time here for example, where others take ages when score drop which >i explain by a program that needs too much overhead). If you have pawn = 32 >then MTD works great of course. > >If you have pawn = 1 it works EVEN better. > >the more silly your evaluation the better. of course major bugs which cause >big deviations in evaluation are not very good. > >However fritz is well tuned. > >But by far best working is just returning material > pawn=1 > knight=bishop=3 > rook=5 > queen=10 > >try it with these piece values. then compare all algorithms. you'll see, >MTD will outperform *anything* with simple evaluations. With simple evaluations (material only) I think you are right. Such experiments will be closer to Aske's ones. Once I worked for some time on the MTD(f); my major problem was the odd/even problem of the evaluation function. Although using the value of two plies ago (instead of last ply) minimized this problem, I still got better results with the PVS. Well, but maybe I didn't spend enough time on that... > >In fact the test has been done already. Ask Rudolf Huber after his MTD >experiences with material only at his celeron processor.
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.