Author: Robert Hyatt
Date: 06:58:59 09/25/01
Go up one level in this thread
On September 25, 2001 at 02:58:22, José Carlos wrote: >On September 24, 2001 at 23:41:17, Dave Gomboc wrote: > >>On September 24, 2001 at 15:24:53, José Carlos wrote: >> >>> First, the context: >>> >>> Last night I finished implementing pondering for the first time. As usual, I >>>chose the easyest way to begin with, until I make sure I understand everything. >>> So I did it this way: after moving, I guess the opponent move (second of the >>>pv), make it, and start thinking. When the opponent's move arrives, I unmake the >>>guessed move, make the real move, and start thinking normally. I expected the >>>program to get to the pondering depth due to the info in the hash table. >>> This worked fine most of the time, but when the ponder search failed low deep, >>>the research didn't go straight to that point. Instead, it chose another move at >>>the begginning (because it saw the bad move in the hash table) and went >>>deepening slowly. >>> I was very disapointed with this behavior, but when I started playing on ICC, >>>I saw a big rating increase. Actually, the explained behaviour turned out to >>>work really good, as usually the program made a good move even with less depth. >>> >>> Now the question: >>> >>> Has this been tried in _normal_ search? I mean, restarting from the begginning >>>after a deep fail low. >>> Is this it a mistake to do what I'm doing? If so, what are the drawbacks? >>> >>> Thanks in advance, >>> >>> José C. >> >>Yes, many years ago Jonathan Schaeffer found this method (restarting the search >>after a big fail-low) worked well for him. Make sure you give yourself enough >>time to get back to a reasonable depth, though (for example, the same depth that >>you found the fail-low at). Many people don't do it, though, so perhaps they >>have found otherwise. >> >>It's funny that you were initially "disappointed"... why is that? Did you think >>that your method was a bit of a hack and was doing the wrong thing in this >>circumstance? :-) >> >>Dave > > I was disappointed because I expected the search to get quickly to the point >where the ponder search finished, given the info in the hash table. I thought my >method was wrong, and I must change it, until I noticed that it was actually >working fine. > > José C. When you think about it, it could not. That deep fail low table entry will re-shape the entire shallow search tree and make you traverse parts of it that you quickly skipped on the first search. But now you are doing a more informed search (you have a piece of information you didn't have the first time) so you should expect that hash hits are not going to get you back to the last iteration quickly.
This page took 0.01 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.