Author: Robert Hyatt
Date: 15:30:05 11/13/99
Go up one level in this thread
On November 13, 1999 at 15:57:31, Bella Freud wrote: >On November 12, 1999 at 11:47:21, Robert Hyatt wrote: > >>On November 11, 1999 at 12:53:35, Ratko V Tomic wrote: >> >>>> I believe that the deeper you go, the more accurate your 'scores' have >>>> to be (by scores I mean weights for each positional thing you recognize). >>> >>>The reason for this is the error propagation effect, which as you propagate the >>>uncertain scores up the tree, increases the uncertainty of the backed up score >>>(unless the leaf uncertainty was 0, such as if checkmate was found, in which >>>case the error remains 0). Altough there are artificial "pathological" games >>>where the error grows exponentially with depth (Judea Pearl's book "Heuristics" >>>has a chapter on this error propagation and gives such examples), in chess, or >>>other "normal" strategy games, the error band behaves as it propagates like a >>>random walk, spreading as the sqrt(Depth). >> >>There is another issue. If you search to shallow depths, you only have to >>choose between A, B and C. One positional 'asset' may be all you can get. >> >>If you search deeper, you have to decide if A+B is better/worse than C, A+C is >>better/worse than B, etc. IE at shallow depths, just recognizing key positional >>features is enough. At deeper depths you have to be able to differentiate >>between various combinations of the positional features.. which makes it a good >>bit harder. >> >>IE an important point (to me) is that a program tuned on hardware that searches >>to depth=D is probably _not_ optimally tuned to run on hardware that can search >>to depth=D+n where n=1,2,... I ran into this many times working on Cray Blitz, >>where we tested/tuned on a vax, and found that the tuning was wrong when we >>moved to the Cray for competition. >> >> > >May I make a comment here? > >You tuned on a vax because it was cheaply available. Those Crays have many other >things to do, and their time doesn't come cheap. > >Therefore, I understand that your program was tuned on a vax. > >Then it was moved to a Cray for competition. It must have been tough to get the >time. > >So with the Cray for the competition only, how would you know that your program >was out of tune? Wouldn't you need to show by then tuning on a Cray that the >tuning was out - by tweaking the tuning. the basic approach was 'deep searches'... on 'key positions'. But on a vax, at 100nps, it took a _long_ time to match the cray at 80K nodes per second (or more). > >But no Cray would have been available for these tuning experiments. > >Were you in fact engaging in the wishful thinking that your program was better >than the shown performance and that the perceived worsefulness was a result of >being 'out of tune'? > >Please forgive my mere observation about the scientific method. > > >Bella what we did was 'tune' on the vax, as best we could. And if we saw something particularly odious (as we did in the 1986 WCCC) then we would adjust whatever we found. In some years, we got more cray time (particularly in the 1980-1986 time-frame). In early years, cray actually gave us time to play in the annual ACM event _plus_ one other 'tune-up' event (ie CB was the first program to ever win a state championship event in 1981, doing just such a 'tune-up'). But such tune-ups became harder and harder to schedule. Some things were easy (ie adding singular extensions). Other things were harder (adding more chess knowledge). This was the main reason CB did worse after the 1986 event. We simply never had a chance to _really_ test the thing, and when we did get machine time, we mainly had to test the parallel search stuff...
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.