Author: Robert Hyatt
Date: 15:17:35 07/27/04
Go up one level in this thread
On July 27, 2004 at 14:55:18, Sune Fischer wrote: >On July 27, 2004 at 13:31:58, Robert Hyatt wrote: > >>>Remember, you generated that target time from a very lose ad hoc algorithm in >>>the first place, not taking anything search related into account. >>> >>>Therefore it is going to be very coincidental how much time you get left over >>>for the final iteration. >>> >> >>Yes, but it is not really "ad hoc". I have N moves left to play, and I have M >>minutes of time to do so. So I really _do_ have to have a plan that (a) lets me >>search long enough to find necessary tactics; (b) saves enough time so that >>should I get into some sort of difficulty before the time control I have time to >>search my way out of it; (c) I have reasonable time left for the moves just >>prior to time control so that I don't make a simple tactical error. It is all >>related.. > >Hmm, I don't see what this has got to do with anything..? It addressed the following statement, quoted from above... ============================================================================== >>>Remember, you generated that target time from a very lose ad hoc algorithm in >>>the first place, not taking anything search related into account. ============================================================================== > >>>>I _always_ start the next iteration even if there is only one second left, as I >>>>don't do my time check in Iterate() but only inside the search itself. >>>> >>> >>>I don't understand your comment. >>> >>>Do you mean that just because you _always_ do something one particular way, it >>>is evidence that it can't be done any better? >> >> >>Nope. I said what _I_ do. Very clearly. And I said what _I_ believe. You can >>always take it with a grain of salt. My time allocation stuff comes from many >>years of watching and tweaking. Not pulled from a hat with no thought... > >Then what is your idea with starting a new iteration when you already _know_ it >can't possibly have time to return a result? I'll say this again. I do _not_ "know" anything. Often a fail-low takes just a few seconds. It is far easier to search than a normal or fail-high root move. So I start a new iteration and give it a chance to fail low. On occasion it does. On occasion it doesn't. Either way there is little cost. After I make a move, I get a running start on the re-start of the search due to the transposition table. > >That's what I'm questioning here, because it can really only return fast enough >if it fails low, but we already have much more efficient ways of checking for >that. What are they? Minimal window? You fail low on the first move with a minimal window. Now what? Continue searching until you get a fail-high and hope you just get one so you know that move is better? Play that? Done in Belle and Cray Blitz circa 1980. Problems. What to ponder? No idea as you failed hi at the root, you failed low on every ply-2 move and you have no idea which is best. The first PVS type search any of us used worked like that. you get a fail high on a non-first move, why expand the window and search again? You only need to do that if a second move fails high as you now need to know which is best. It works. But I did that (as did Belle) and suffered from a lack of something good to ponder as that begins to happen a lot when you code the null-window search that way. Not bad. But I wanted a real score, a real move to ponder, etc. > >>What your question has to do with my statement about not checking the time at >>the end of an iteration, but only inside the search, I have no idea.. > >I never said anything about time inside the search. You posted your comment immediately after my comment about exactly that, however... > >I'm talking about what to do at the last moment before exitting, when time is >really sparse and we can say "for sure" we don't have enough time to resolve a >new score. > >-S. How can we say that "for sure"? The tree is way beyond variable.
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.