Author: Robert Hyatt
Date: 10:18:13 11/14/00
Go up one level in this thread
On November 13, 2000 at 22:48:23, Uri Blass wrote: >On November 13, 2000 at 15:26:15, Robert Hyatt wrote: > >>On November 13, 2000 at 13:58:49, Uri Blass wrote: >> >>>On November 13, 2000 at 11:55:08, Robert Hyatt wrote: >>> >>>>On November 12, 2000 at 15:10:43, Uri Blass wrote: >>>> >>>>>On November 12, 2000 at 13:25:15, Robert Hyatt wrote: >>>>> >>>>>>On November 12, 2000 at 12:48:14, Uri Blass wrote: >>>>>> >>>>>>>On November 12, 2000 at 11:05:30, Robert Hyatt wrote: >>>>>>> >>>>>>>>On November 12, 2000 at 10:54:42, Jeff Lischer wrote: >>>>>>>> >>>>>>>>>It seems if you correctly predict the opponent's move 100% of the time, this >>>>>>>>>would correspond to doubling your available time (you would be thinking on your >>>>>>>>>time as well as your opponent's time). If a doubling of speed results in an Elo >>>>>>>>>improvement of 60-70 points, is this also the maximum benefit for permanent >>>>>>>>>brain? With diminishing improvements at longer time controls, the benefit might >>>>>>>>>be even less? >>>>>>>>> >>>>>>>>>If the above is correct, then what about the case where you correctly ponder >>>>>>>>>only 60% of the time. This seems like a pretty typical value. Then is the >>>>>>>>>benefit only about 40 Elo points? >>>>>>>>> >>>>>>>>>Are there any other approaches to permanent brain that might be more effective? >>>>>>>>>At first I was wondering about simply searching on your opponent's time like you >>>>>>>>>do on your turn -- using selective searching to focus on the best moves. But >>>>>>>>>then I thought of another possibility. What about a different kind of searching? >>>>>>>>>Maybe search using lots of knowledge during your opponents time trying to >>>>>>>>>develop a plan? Or maybe do a fast selective search looking for killer tactical >>>>>>>>>shots? >>>>>>>>> >>>>>>>>>Humans think differently on their time versus their opponent's time. Maybe >>>>>>>>>computers would benefit from doing the same? I don't know enough about chess >>>>>>>>>programming, however, to know how (or even _if_) the results of that "opponent's >>>>>>>>>time search" could get passed to the "your time search". Would hash tables be >>>>>>>>>sufficient? >>>>>>>> >>>>>>>> >>>>>>>>This has been answered before... here is the quick version of the idea: >>>>>>>> >>>>>>>>let's take two different pondering algorithms: (1) present idea where we >>>>>>>>assume that the best move from the last search is searched for the entire >>>>>>>>time; (2) alternative where the best N moves are searched (less deeply of >>>>>>>>course). >>>>>>>> >>>>>>>>case 1: target search time is 3 minutes. The opponent takes three minutes >>>>>>>>to make his move. >>>>>>> >>>>>>>This assumption is not correct. >>>>>>>The opponent(espacially in cases that the opponent is human) may use 30 minutes >>>>>>>for one move and less time for the other moves) >>>>>>> >>>>>>>I believe that in this case it is better to stop searching the best move after >>>>>>>part of this time and start to consider the response for the second best move. >>>>>>> >>>>>>>Uri >>>>>> >>>>>>The same thinking applies. I am _sure_ I am going to predict his move over >>>>>>50% of the time. If he takes a long time, should I take a long time, or should >>>>>>I do a bunch of three minute searches on different moves he might choose, and >>>>>>after _his_ long think I play a move found after a 3 minute think? >>>>>> >>>>>>I think the current approach is best for _all_ circumstances... >>>>> >>>>>It is not clear. >>>>> >>>>>The benefit that you earn from another 3 minutes of search after you search for >>>>>more time is smaller because of diminishing return from speed and in cases that >>>>>you did not predict the move correctly in the first try you can earn the first 3 >>>>>minutes that are more important. >>>> >>>>Note that I don't necessarily agree with the concept of "diminishing returns" >>>>when it comes to search depth. >>> >>>It is clear that there is also diminishing return in search depth. >>> >>>The average search depth that you get from another minute is lower when you use >>>more time so if you decide to start analyzing a reply for another move after 9 >>>minutes of pondering you sacrifice less plies for the same target relative to >>>the case that you do it after 3 minutes and it is logical to think that at some >>>point the gain may be bigger than the sacrifice. >>> >>>Uri >> >> >>I am not sure what you mean. When I ran the crafty goes deep tests, the >>search depth was pretty linear. for every factor of 3x or so more time, it >>went one more ply deeper. I didn't see any "wall" that it ran into. It does >>help if you run it with a huge hash for huge searches. > >I assume the same branching factor(let assume it is 3). > >The point is that if you search the reply to the predicted move for 9 minutes >instead of 12 minutes then you lose only (log(12/9))/log(3)) plies >so you lose only (log(12/9)/log(3)) plies for another 3 minutes search that can >help you in case that the opponent does not play the predicted move. > >If you search the reply for the predicted move for only 3 minutes you will lose >(log(6/3)/log(3)) plies that is more plies for the same gain. > >I do not know when is the right point to start searching for reply for another >move but it is logical to assume that at some point the gain is bigger than the >loss. > >Uri Here is the main point: I _know_ that I am going to predict correctly over 50% of the time. Which means I _know_ that if I search a second move, I am going to be wasting that time in most of the cases... I have seen cases where an extra 10 seconds was enough to fail low, go into a "panic time" (to use the DB term) search and find a way out. This can happen at deep search depths as well as at normal depths. But in any case, if I know that something happens more than 50% of the time, then preparing for the <50% case doesn't seem to be a good use of resources. In the Kasparov/Kramnik matches, for example, Crafty's prediction rate was well over 75%. In one game it was over 90%. It is hard to imagine that spending time on a move likely to be wrong at least 75% of the time is a good investment.
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.