Author: José Carlos
Date: 00:16:18 02/09/00
Go up one level in this thread
On February 08, 2000 at 19:04:59, John Stanback wrote: >On February 08, 2000 at 18:24:41, José Carlos wrote: > >>On February 08, 2000 at 14:28:17, John Stanback wrote: >> >>>On February 08, 2000 at 10:09:27, William Bryant wrote: >>> >>>>I need some recomendations on handling time management when the program >>>>fails low at the root near the end of the alloted search time. >>>>(During game play, not during test suites) >>>> >>>>How do people handle this? >>>> >>>>Do you research the PV move at the lower window (which will return a >>>>move and a score) and then stop? >>>> >>>>Or, do you need to search all the root moves at the lower score to determine >>>>which is truly the best move? >>>> >>>>The later makes more sense, but may consume a significant chunk of additional >>>>time which the program may or may not have. >>>> >>>>Thoughts, ideas and suggestions gratefully accepted. >>>> >>>>William >>>>wbryant@ix.netcom.com >>> >>>If the first move fails low, I increase the time allotment by 10X >>>and then search the remaining moves using the same window, but with >>>a reduced depth. >> >> I don't understand researching with reduced depth. Won't those moves, with the >>reduced depth, be in the hash table from previuos iterations? >> >> José C. >> > >Say I'm searching at depth=10 and the 1st move fails low. Then I >continue to search all the moves in the list with the same window, >but I reduce the search depth to 8. You're right, these moves should >be in the hash table, probably with the score being an upper bound >since they were not the best move at previous iterations. If the >score is below the current alpha value and is an upper bound then >that move can be ignored. But if the score is above the current alpha >value then it still needs to be searched. If the 8 ply search comes >back with a score above alpha, then I immediately search it again >to the full 10 ply. > >The reason behind this technique is that sometimes the first move >fails low and there are one or more other decent moves. I think this >method finds these cases faster than trying to get an accurate score >for the first move before searching the rest. If there aren't any >other decent moves, then often I don't waste too much time doing the full >10 ply search. Thanks for the explanation. I'll think about it :) José C. >> If any of the remaining moves don't fail low at >>>the reduced depth then I immediately re-search the move to the correct >>>depth. If I get a move with an ok score (not fail low) at the >>>correct depth and the original search time limit is exceeded then >>>I play that move. If all moves are searched this way and fail low >>>then the window is lowered and the search procedure starts over. >>> >>>By the way, I use windowed alpha-beta rather than PVS. >>> >>>John
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.