Author: Robert Hyatt
Date: 08:12:12 09/27/01
Go up one level in this thread
On September 27, 2001 at 10:31:04, Uri Blass wrote: >On September 27, 2001 at 09:39:22, Robert Hyatt wrote: > >>On September 27, 2001 at 03:59:57, Uri Blass wrote: >> >>>On September 26, 2001 at 21:45:46, Robert Hyatt wrote: >>> >>><snipped> >>>>Here is the problem.. I had to explain this to Komputer Korner a few years ago >>>>as well... >>>> >>>>If you correctly predict your opponent's move at least 50% of the time, or >>>>more, then the way we currently ponder can _not_ be improved on. >>> >>>Not correct. >>>There are ways to improve the pondering but not the way that bdann suggested. >>> >>>For example if your program ponders for the opponent move and fail high and find >>>that it is winning after the fail high then it is possible that the program >>>expect a blunder and it may be better to go back to the root position to see if >>>the opponent has a better move and to ponder for the better move. >> >> >>So? If you are predicting > 50% of the time, then there is not a thing wrong >>with sticking with this. It can fail high this iteration, then fail low the >>next one. > >I did not say after every fail high but only after you see that you are winning. >it depends on the score. Have you never seen a program fail high, get a score of +3, which should be clearly winning, only to fail low in the next iteration? I saw this in the last WCCC event where one program was at +5 and _lost_. I don't see how you can say "this is good enough, let's move on to another move." Because a move that can fail high can also fail low just as easily on a deeper search, in very sharp tactical positions. > >An extreme example is when you see a forced mate after the expected move. >In this case you can safely stop pondering on the opponent move and ponder on >another move. >If you see a score of +20 pawns for yourself you also can go to ponder for >another move. That doesn't even happen once per game, so it is a non-issue from a statistical perspective. > >I am less sure if it is a good idea to ponder on another move when you see a >score of +5 but if you can be sure in 99.9% that you win after seeing +5 score >then it can also be a good idea because taking a risk in 0.1% of the cases is >not a big risk and the risk that the opponent does not play the predicted move >is bigger. no it isn't. If your opponent plays the predicted move, you get a long think to make sure your tactics are correct. If you spread the time over more than one move, then you search less deeply. But no matter what, if you don't predict correctly, you will _still_ think for the normal amount of time, anyway. >> >> >>> >>>Another idea is in cases that the opponent goes for a long think(for example 30 >>>minutes at 40/2 hours game) for one move and in this case it may be better to >>>stop pondering after the program pondered for enough time and to start pondering >>>for the next best move or for the root position. >> >> >>Again, this doesn't work. Given the choice of pondering a move I believe will >>be correct over 50% of the time, vs pondering two different moves for 1/2 of >>the time, I'll take the former _every_ time. And get my _own_ deep search in >>since my opponent thinks something is worth a deep think also. So long as I >>predict > 50% of the time there is _no_ better methodology. > >The difference between 3 minutes of search and 6 minutes of search is more than >1/2 ply assuming a branching factor of 3(log6/log3 plies). Depends. I consider 1/2 ply to be searching just the PV. Not the rest of the moves. > >The difference between 30 minutes of search and 33 minutes of search is >smaller(log33/log30 plies). > >sacrificing log6/log3 plies for searching another move may be a bad idea but >sacrificing only log33/log30 plies for the same target is a better deal. > >Uri
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.