Author: Vincent Diepeveen
Date: 04:59:55 04/15/02
Go up one level in this thread
On April 14, 2002 at 16:24:44, Alessandro Damiani wrote: >On April 14, 2002 at 13:57:38, Vincent Diepeveen wrote: > >>On April 14, 2002 at 13:37:19, Alessandro Damiani wrote: >> >>>Hi Vincent, >>> >>>You too much concentrated on the game Mawari. I think they choose Mawari to have >>>a simple framework to experiment with. I guess a Mawari engine is far simpler >>>than a Chess one. So, forget Mawari. :) >>> >>>You are right, alpha-beta evaluation functions are like lazy evaluation. But, >>>the big difference is that an alpha-beta evaluation function is an algorithm >>>that traverses a classification tree. I have in mind the picture of an ordered >>>hierarchical structure of position features (a tree of features). At first sight >>>it seemed to me like that (right, I didn't take the time to read the whole text >>>:). >>> >>>We both agree on the bad effect of lazy evaluation on positional play, but an >>>alpha-beta evaluation function seems to be different: the bounds on a feature's >>>value range are not estimated. >>> >>>But maybe I am wrong. >> >>Yes you are wrong. >> >>Let them show pseudo code. >> >>Then you see what they describe is 100% the same and completely >>not working. >> >>What they do is kind of: >> >> "we have a great new feature and we call it X" >> >>In reality they invented old feature F which was already proven >>to be not working. >> >>So in fact 2 mistakes are made by them >> a) not checking out existing literature and experiments done >> b) committing scientific fraude by describing something >> existing. >> > >I understand what you mean, but it is better if you first have got more >information before you judge. Here is the pseudo code taken from the text: So if a position is a leaf it always returns k.alpha, or in short you *never* get a value for your search. Only innernodes might get evaluated. Sorry no one can use this. apart from that, suppose the line is not there to return k.alpha when it is a leaf. Then let's take DIEP. I have a few thousands of values which can apply to a single position. If i just evaluate half of them i always end up with like 30 pawns. Not possible to just evaluate right or left! >function static_evaluation(p: position; > alpha, beta: real; > k: evaluation node): real; >begin > for i:= 1 to D do unknown[i]:= true; > > while true do begin > if k.beta <= alpha then return alpha; > if k.alpha >= beta then return beta; > if leaf(k) then return k.alpha; > > if unknown[k.feature] then begin > vector[k.feature]:= get_feature(p, k.feature); > unknown[k.feature]:= false > end; > > if vector[k.feature] <= k.split_value then > k:= k.left > else > k:= k.right > end >end > >where D is the number of features in a position. > >Here is the link where I took the text from: > > http://satirist.org/learn-game/lists/papers.html > >Look for "Bootstrap learning of alpha-beta-evaluation functions (1993, 5 >pages)". > >Alessandro > >>> >>> >>>On April 14, 2002 at 11:42:34, Vincent Diepeveen wrote: >>> >>>>On April 14, 2002 at 04:26:52, Alessandro Damiani wrote: >>>> >>>>Seems to me that these idiots never have figured out what already >>>>has been tried in computerchess world. >>>> >>>>Of course i'm not using their 'concept' which already exists >>>>by the way. These guys are beginners everywhere of course. >>>>Mawari, every idiot who programs for that game can get >>>>world champ there of course, or pay levy to get a gold medal... >>>>...if i may ask... >>>> >>>>What works for a 2000 rated chessprogram to experiment >>>>with doesn't work for todays strong chessprograms simply. >>>>Mawari programs when compared to chess programs are at 2000 >>>>level of course, relatively seen to how much time and >>>>effort has been invested in mawari programs. >>>> >>>>If i read their abstract well then in fact they define a >>>>'partial' evaluation, already known under the name >>>>lazy evaluation using a quick evaluation. >>>> >>>>That's a complete nonsense approach. It's pretty much the same like >>>>lazy evaluation based upon a quick evaluation, it's most likely >>>>exactly the same, if not 100% similar. >>>> >>>>If i would describe here how much time i invested in making >>>>a quick evaluation which evaluates some rude scores, and which >>>>with some tuning when to use it and when to not use it, that >>>>it always scores when used within 3 pawns in 99% of the positions, >>>>then people would not get happy. >>>> >>>>I invested *loads* of time there in the past. >>>> >>>>More important, i generated big testcomparisions here to see >>>>when the quick eval worked and when not. That's why i could >>>>conclude it didn't work. >>>> >>>>Even more unhappy i was when i tested with this concept. Disaster. >>>>Yes it was faster concept, but here the amazing results >>>> - positional weaker >>>> - tactical weaker >>>> >>>>the first i wasn't amazed about of course, but the second i was. >>>>i was pretty amazed to find out that these 1% of the evaluations >>>>where the quick evaluation gave a score but evaluated it wrong, >>>>really amazing that these evaluations cause a tactical way better >>>>engine. >>>> >>>>Simply majority of tactical testset positions get solved by evaluation >>>>and NOT by seeing a bit more tactics. >>>> >>>>In short it's not working simply to use a lazy evaluation in a program with >>>>a good evaluation which also has high scores for things like king >>>>safety. >>>> >>>>>Hi all, >>>>> >>>>>I am wondering if someone uses "alpha-beta-Evaluation Functions" by Alois P. >>>>>Heinz and Christophe Hense. Below is the abstract of the text. >>>>> >>>>>Alessandro >>>>> >>>>> >>>>>Bootstrap Learning of alpha-beta-Evaluation Functions >>>>>Alois P. Heinz Christoph Hense >>>>>Institut für Informatik, Universität Freiburg, 79104 Freiburg, Germany >>>>>heinz@informatik.unifreiburg.de >>>>> >>>>>Abstract >>>>>We propose alpha-beta-evaluation functions that can be used >>>>>in gameplaying programs as a substitute for the traditional >>>>>static evaluation functions without loss of functionality. >>>>>The main advantage of an alpha-beta-evaluation function is that >>>>>it can be implemented with a much lower time complexity >>>>>than the traditional counterpart and so provides a signifi >>>>>cant speedup for the evaluation of any game position which >>>>>eventually results in better play. We describe an implemen >>>>>tation of the alpha-beta-evaluation function using a modification >>>>>of the classical classification and regression trees and show >>>>>that a typical call to this function involves the computation >>>>>of only a small subset of all features that may be used to >>>>>describe a game position. We show that an iterative boot >>>>>strap process can be used to learn alpha-beta-evaluation functions >>>>>efficiently and describe some of the experience we made >>>>>with this new approach applied to a game called malawi.
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.