Author: Robert Hyatt
Date: 08:40:08 07/11/02
Go up one level in this thread
On July 11, 2002 at 00:34:56, Ricardo Gibert wrote: >On July 10, 2002 at 22:04:44, Robert Hyatt wrote: > >>On July 10, 2002 at 20:42:45, Ricardo Gibert wrote: >> >>>On July 10, 2002 at 11:04:34, Sune Fischer wrote: >>> >>>>On July 10, 2002 at 09:30:30, Ricardo Gibert wrote: >>>> >>>>>On July 10, 2002 at 04:30:28, Sune Fischer wrote: >>>>> >>>>>>On July 10, 2002 at 01:02:38, Ricardo Gibert wrote: >>>>>> >>>>>>>I find it fascinating that so many very experienced computer chess programmers >>>>>>>do not understand some rather fundamental properties of alpha-beta search. >>>>>>> >>>>>>>What RH ran into was actually quite predictable. As MF put it, "I'm surprised >>>>>>>that you're surprised." >>>>>> >>>>>>In a PV search some of the nodes are searched twice, so if you get something >>>>>>outside the zero window because of a collision, it slows you down as you have to >>>>>>research, but it doesn't hurt you unless you get another collision while >>>>>>researching. >>>>>> >>>>>>-S. >>>>> >>>>>This is not a property of using PVS. A normal alpha-beta search plus hash table >>>>>does this too. >>>> >>>>How so? >>>>I would say it does the opposite, since its about refuting moves. >>> >>>What either one of us has to say on this is really irrelevant. RH needs to >>>conduct a more reasonable test. The manner in which he generated the collisions >>>plus the use of PVS should be changed. >> >>How would you suggest generating the collisions? As far as dropping PVS >>however, I disagree on. That is what I _use_ in real games. That is what >>I want to understand the behavior of. I don't care about variants of >>alpha/beta that I am not using myself (ie mtd(f) or something else that >>is non-PVS). > >You are forcing a collision every 1000 nodes, but then the same position will >not likely produce the same collision when the position is researched. That is true in PVS regardless. The first false match is almost guaranteed to be a >beta or <alpha type match. On the re-search, this will be useless _anyway_. > You are >testing something very different and valuable than if you were generating the >collisions in a consistent way unless I have misunderstood what you were doing. > I don't think it matters with PVS since the first search is on a null-window but the second (research) is not. There are hardly any "exact" scores in the table anyway, so about all we will see are fail high/fail low entries. And they will be useless on any re-search. >I think studying alpha-beta first will afford a more precise understanding of >what is going on. Start with the simple before moving on to the more complex. I >just think it is better to be more methodical. My "first cut" was not intended to be a 100% determining test. It was intended to get some initial data, without requiring a huge amount of work. The first cut added 3-4 lines of code. Removing PVS would be more complicated than that... I'm not sure I will ever care what pure alpha/beta does since I never plan on using it again, but futher experiments will certainly be more comprehensive than this "pilot study". > >> >> >> >> >>> >>>> >>>>But thinking further about it, the collision could also return a false score >>>>outside alpha-beta, so the move is refuted when it should not be. I don't know >>>>which is worse, but at least landing _inside_ is harmless as we just research, >>>>which was my point. >>>> >>>>-S.
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.