Author: Daniel Shawul
Date: 00:44:06 08/27/05
Go up one level in this thread
On August 26, 2005 at 19:00:26, Volker Böhm wrote: >On August 26, 2005 at 14:21:34, Alvaro Jose Povoa Cardoso wrote: > >>Hi, >>some of you compare the number of times a move failed high to the number o times >>the same move failed low in order to decide if a move can be reduced one ply. >>I've tested this and also tested using the actual values of the history table >>(using of course another history table for fail lows). >>I couldn't reach a conclusion though. >>What is your experience on this? >> >>best regards, >>Alvaro > >Hi Alvaro, > >the first time I tried history based pruning I was allmost sure that this >couldn´t work. But now it works fine for me. One point I have to add is that I >only count a fail low if there is some move sorted after it that failed high, >thus I do not count fails on alpha nodes. > >I can´t really beleive that there are really many moves that are more often >failing high than failing low. But it is the case. > >There are many parameters you can experiment with: > >1. how many points you add to a history-pruning table for a move failing >high/low dependant on the current search depth. (example square of depth?) >2. Which moves are not reduced even if the fail high/fail low ratio is "bad" >(example: captures, check moves, promotions, ...) >3. At what ratio do you reduce and how many plies do you reduce depending on the >current search depth >4. What do you do if a reduced move fails high (search again without reduction?) >5. "Hysterese": Prevent a toggling of pruning/non pruning decisions by adding >some points if a move switches from pruning to non pruning. >6. Interaction between moves and pieces (example a rook move that leads to a >double rook on a column could be handled different as the same rook move without >doubling rooks). >I have tried at least 50 different combinations of those parameters. For me I >have found one combination that improved strength much. But the combination is >far away from the combination I had expected to be best. > >The only advice I can give for this type of pruning: don´t think to much, just >try! And keep it simple! > >Greetings Volker Glad to hear that you tried many combos. I think that the reason why "history pruning" works is not because the history data is perfect but just becasue of the great nominal search depth it brings. If you make it non-recursive, depth will not be that good, and you will immediately feel the weaknesses which were hidden by search depth. Reducing moves that are likely not to cause a fail high based on some criteria ,be it history or not, seems to work well. i don't think history data is the only thing which works. My reduction criteria is , for example, don't reduce killer moves and the first n non-capture moves. This seems to work well. I also tried reducing even extended moves. At first this seemed to work because it gets even more depth. i disabled this now because of some tactical problems i can't fix. i think that doing intense math on the history data is not that smart. It is just a random data that gets even more random as search depth increases. Daniel
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.