Author: Robert Hyatt
Date: 11:58:37 04/08/04
Go up one level in this thread
On April 08, 2004 at 14:00:50, Eric Oldre wrote: >When people talk about there move ordering hitting 90%, what exactly are they >measuring? is it that on 90% of the nodes the first move tried is the best? For me, it means that on any node where I fail high, 90% of the time (actually a bit higher for me, more like 92-93% most of the time) I will fail high on the first move I search... > >what i've been doing is averaging the percentage of moves that i go through >before i find the best move. > >example: if there are 20 moves from a position, and the 2nd move evaluated is >best, i'd say i had 5%. (or 95%). if the first move is the best, then i have 0% >(or 100%) > >i store the statistics separately depending on whether it's a pv or beta node. > >for beta nodes i'm getting a move that causes the cutoff very fast, on average >after around 2-3 % of the moves. > >for PV nodes, i't a bit worse, i don't have the numbers with me, bug it seems if >i remember right, about 20%. (or 80%) > >these seems to make since to me since if the parent node made a big blunder, it >shouldn't take long to find something that will cause a cut off, but it the >parent node was a good move, then > >so my question is, am i using the "standard" way of measuring move ordering? and >if not, what is the standard way. so that i can compare apples to apples when >talking about move ordering. I don't know if there is a "standard way" at all. I have been reporting using "my way" for a long while and it is a simple but easy to understand idea. IE in SMP search, it doesn't matter whether a PV or non-PV move has bad move ordering, all that matters is that you parallel search at a node where all branches need to be searched. The higher "my" percentage, the more likely you are to make a good parallel search split decision... > >Thanks! > >Eric Oldre > >On April 08, 2004 at 11:50:31, Robert Hyatt wrote: > >>On April 08, 2004 at 09:08:19, Andrew Wagner wrote: >> >>>Hi all, >>> >>>I'm using a simple hashing scheme: a single replace-always tranposition table. >>>I've made sure all the hash keys are correct by comparing it to a key created >>>from scratch. But I'm not sure it's working as efficiently as it could be. So, >>>my question is: what statistics can I generate that will tell me how it's doing? >>>And what values should I be getting for those statisics, on average? >>> >>>Also, a slightly different topic: someone told me that with hash tables, and >>>without null-move, my move ordering (first-move fail-highs / total fail-highs) >>>should be averaging >95%. Does that sound right to the rest of you? Thanks! >>>Andrew >> >> >>More like > 90%, with _some_ hitting >95%...
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.