Author: Robert Hyatt
Date: 12:38:01 11/23/01
Go up one level in this thread
On November 23, 2001 at 14:31:39, Uri Blass wrote: >On November 23, 2001 at 10:57:08, Robert Hyatt wrote: > >>On November 23, 2001 at 03:45:35, Uri Blass wrote: >> >>>On November 22, 2001 at 22:50:31, Robert Hyatt wrote: >>> >>>>On November 22, 2001 at 21:26:37, Dan Andersson wrote: >>>> >>>>>>It is anectodal from the perspective I gave... namely that of playing >>>>>>A vs A (different depths) to extrapolate how A does at increasing depths >>>>>>against _anybody_. >>>>>> >>>>>I have to agree, but in its own context it would be called substantiated. The >>>>>context or contexts of the different points of your posting was, IMO ambiguous. >>>>>That's why i posted the factoid in return. I hope someone with more time in hand >>>>>makes a similar test, in regards to accuracy and reliability. Many obstacles to >>>>>generalisation to A B matches when A neq B exist. The internal definition of ply >>>>>is one. Different extensions strategies is another. Bugs that occur rarely but >>>>>wastes good play, or are a function of depth being a prime or generally a >>>>>function of depth ... etc. It might even be that diminished returns between two >>>>>different programs is dependent on too many factors to be measured reliably. One >>>>>criterion that ought to be fulfilled before trying to find diminishing returns >>>>>between two different programs. One that needs to be there, is that both >>>>>programs show diminishing return in self play testing. I cannot give a valid >>>>>reason right now. But I have a hunch that it might be an almost necessary >>>>>prerequisite. Any thoughts? >>>>> >>>>>MvH Dan Andersson >>>> >>>> >>>>Something tells me that for A vs A, there _must_ be a diminishing return, >>>>because all that changes is the depth. But in A vs B, the search depth is just >>>>one difference between the two players. >>>> >>>>I _always_ find positions where another ply (or another N plies) would find >>>>the right move... >>> >>>I do not understand why do you think that in A vs A there must be a diminishing >>>returns. >>> >>>Is it the situation in all the other games? >>> >>>Uri >> >>The math suggests it is true. If the _only_ thing that is different between >>player (A) and player (B) is one ply of search, then going from 19-20 is less >>of a change than going from 4-5. It is possible that even this is not true, >>of course, but intuitively it should be so. > >I do not see a reason to assume that intuitively there is diminishing return if >you play the program against itself when you have not diminishing returns when >you play with different programs. > >Suppose that you have a game when half of the positions A does not understand >and half of the positions B does not understand. > >At small depthes tactics is going to dominate so more plies are going to help >between A and B. > >At big depthes in half of the cases A is going to lose the positions that it >does not understand and in the second half B is going to lose the positions that >it does not understand and you are going to get 50%. > >The depth is not going to be important because at every depth that is smaller >than 30 plies and bigger than 10 plies the program that does not understand the >position is not going to have enough depth to solve the problem by search. > >Uri I can envision programs with an evaluation that gets better with depth, which would make them behave _differently_ against other programs as opposed to how they would self-play. Whether this really happens or not is only a guess at present. But concluding that diminishing returns happens _only_ by playing A vs A is also nothing more than a guess when extrapolated to the general case of all programs at deeper depths.
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.