Author: martin fierz
Date: 22:40:41 08/20/02
Go up one level in this thread
first, just to get this out of the way: i think this is quite a valid master's thesis. a MS is by definition not new, original work. that's what a PhD is for. however, i am slightly astonished by the claims: > It gives quantitative evidence for improved hash table usage for n-probe >(n>2), has this really never been published? i made the exact same experiment with my checkers program years ago... i can't really imagine that this is new. what is missing in his thesis is the identification of the optimal probe number - making 8 hashprobes slows the program down, compared to only 2, for example. for me, 8 was gaining too little in reduced nodes compared to the speed cost. comparing node counts is IMO comparing the wrong thing - he should be comparing the time to run through his test suite at a certain value of n. (BTW, this makes a good example for a parameter that you can adjust depending on how deep you search/how powerful your computer is. christophe theron has claimed here that a good program performs equally well on any level; and that optimizing for a certain processor speed is not possible (talking about palm / PC - big difference!)) >it casts serious, quantative, doubt on the effect of the counter-move >technique in combination with domain-smart (SEE) move ordering, i've never heard of the counter-move thing before, so i'm not really surprised that it does no good. another point here: when combining different move-ordering techniques (i use hash/killer/history/static eval in my checkers program), you can choose the weights of every heuristic. obviously the overall effect depends quite a bit on these weights, and therefore this should be examined too. in his example, he is weighing the counter-move higher than a SEE move. which turns out to be a bad idea, according to his data. but the counter-move concept may still be sound and improve node counts if weighed less than SEE. as long as this is not tested, he cannot claim anything about the counter-move technique. this is an omission which should have been caught, by himself or by his supervisors! >rebutes the hash-miss approximation >models that have been pestering academic publications before (even in PhD's). i don't know what exactly has been around, but my back-of-the-envelope calculation i did on hash collisions 2 years ago gave the same result :-) finally, i have a general remark: according to the thesis rookie achieved a 2180 rating in blitz games running on a PII 360MHz (i just looked up on ICC, there it has a blitz rating of 2496 - maybe the 2180 is a fics rating?). still, this is something like 500 elo weaker than really good programs - and that in itself is a hint that whatever he writes about computer chess may be clear, but maybe not very good. aloha martin
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.