Author: Mike Curtis
Date: 21:50:01 05/24/00
Go up one level in this thread
On May 24, 2000 at 10:28:57, Oliver Roese wrote: >Hi all! > >This question is about pondering during the opponents time... >... >It predicts the oppononts move, assuming "optimal" play and then starts to >work until the opponent moves. >If it predicts the opponents move correctly it has a great edge, otherwise >only some hashtableentries. >If it wouldnt predict the opponents move it would gain a small contribution to >_every_ move. >Obviously the better it predicts the opponents move, the better is the first >method. >From my experience as a mere chessplayer i would say the following: >-Predicting the opponents move is very difficult even in games of the >highest value (disregarding trivial cases and extraordinary circumstances). >-Intuitively i would judge a small contribution to every move as more >worthfully than an extremly big one that occurs seldomly >To say it exaggerated: If you have 20 moves to made and distribute >your resources evenly, you may have a chance. If you invest all in the first >move, making the other 19 moves very bad, you are dead for sure. >In more general terms: >The relative benefit of predicted moves decreases rapidly with increasing >searchdepth, i think. > >Maybe one could use a hybrid approach? I think this idea is worth testing. (PVS-Negascout-Mulling, instead of alpha-beta-Pondering) I might use the following approach: 1) Run a performance test suite using standard alpha-beta-pondering. 2) Modify pondering to always halt and restart search, instead of resuming search on ponder hits. 3) Rerun the test suite on the pondering program. 4) Compare the results from 1) and 3). 5) Stare at the screen <g> and decide whether I still believe the performance difference from 4) will be recouped from Mulling on Ponder misses. This is a tough call because Pondering the "wrong" move also fills the hash tables with valuable info, especially if the Pondered move is really the best move. 6) Replace Pondering with Mulling and test again. 7) If the results from 1) and 6) are close enough, (another stare-at-the-screen decision) tweak the program further. For example, Ponder on evading checks, captures, and pawn promotions and Mull on other moves. 8) If this wins, I would try this hybrid approach on the search itself. Before someone chews my head off for being "obviously wrong", Let be humbly admit that I have unleashed neither pondering nor hash tables on the chess servers. :-)) FathomEngine
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.