Author: Eelco de Groot
Date: 19:02:14 12/15/05
Go up one level in this thread
On December 15, 2005 at 06:27:59, Gian-Carlo Pascutto wrote: >On December 15, 2005 at 02:35:39, Chrilly Donninger wrote: > >>Its obvious that the evaluation influences the style of a programm. But >>interestingly the shape of the search tree has also a significant influence. As >>a rule of thumb: The programm prefers lines with larger (sub-)trees. If one >>increases a certain extension the chances for such moves to be played increases >>(to opposite holds for pruning). > >I would say it depends on whether one is winning or losing (score rising or >dropping). When one is winning, the longer lines (extended ones) have generally >better scores, so they get preference. > >When one is losing, the preference is for the shorter lines (the pruned ones). > >This effect is the opposite of what the programmer wants. > That would be one big reason to be cautious not to use too much extensions then; their value depends on the distance to the root. Some sort of new horizon effect, pushing the search over the horizon with extensions, you now fall off the edge running downhill or fly into the sky running uphill :) But I'm sure most programs do to some degree correct for this effect? In practice, when a new move threatens to become best, make sure both candidate PVs are searched to the same depth for instance? Is this too simple a way to view the question? >>I am certain, that this effect exists, but I do not exactly know the reason. One >>explanation is: The evaluation consists of a true term which properly reflects >>the value of the position and white noise. If the programm has many choices, >>because the search tree is large, the expected value of the white noise is >>higher. > >Correct is: the expected extreme values are higher. > >>For non statisticans: If one picks from a bag with 1000 numbers randomly >>1 number and the next time 10 numbers, the chances are very high, that the >>maximum of the 10 numbers is greater than the number picked first. > >It also applies to the minimum (see above), so I don't think this explanation is >all of it. > >-- >GCP I haven't read the whole thread but just my first -non programmers- reaction to this would be to ask whether the minima really matter here? Don't all the low numbers get minimaxed? Another question I had to Chrilly's hypothesis and I haven't read anybody ask this yet, what do we understand as white noise here? Part of it is inaccuracy in the evals, I would agree about that of course but is it not just as simple to assume that if you search longer, you are bound to find more? (More tactics) "Zoek en gij zult vinden!" as we say over here in Holland, and I mean that with a large search tree there would just be more of a chance to find tactical possibilities, which up to a point are also just noise; not every good move increases the possibilities for tactics right away but some will do at some point in the tree, not all tactical gains are good in the long run, but many programs are still materialistic enough to choose them if they arise. But this explanation would be more about comparing a big and a small searchtree, but of the same shape. And the big tree has more tactical noise in it that partly inadvertantly can end up in the PV. But do tall trees have more tactical noise than shallow but broad trees of the same overall size? Is another question. I think they naturally tend to be more tactical, I'll describe below, which as a byproduct could also mean they have more tactical noise. Very probably...(just take another look at your average long overnight search PV, if it did not come mutilated out of the hashbag :) But that sure gets complicated, first I'll give my simple hypothesis about the more tactics. Okay Chrilly was talking about the shape, not the size: my hypothesis would be the longer lines tend to be forced more. They have a longer distance from the root. The higher the depth, the longer this distance from the root, the more sure you want to be that any advantage at the end is tangible (i.e. more material in most cases, especially if you eval is finite in size and accuracy). If the advantage is still just positional you would have had this positional advantage already at the root so then you have made no progress. The purpose is to make advantages more permanent by good play. Therefore a deep searching program tends to be more materialistic (by tuning if not totally on purpose?) and that certainly is of influence on the playing style. Towards a more tactical style probably.. Would be my five cents for the moment! Regards, Eelco
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.