Author: Peter Fendrich
Date: 08:25:01 03/24/04
Go up one level in this thread
On March 24, 2004 at 11:11:02, rasjid chan wrote: >On March 24, 2004 at 07:25:09, Peter Fendrich wrote: > >>On March 24, 2004 at 05:39:20, Tord Romstad wrote: >> >>>On March 24, 2004 at 05:17:56, Peter Fendrich wrote: >>> >>>>Uri didn't invent ETC if that's what you imply! >>>> >>>>Given your story about costly move/unmove functions it's possible that ETC gives >>>>you some savings. Without ETC you will hit the cutoff anyway in the child node >>>>and with smaller unmove costs ETC is not that effective IMHO. >>> >>>It seems to me that you miss part of the idea of ETC. You are right that >>>you will get the cutoff in the child node even without ETC, but in which >>>child node? If your move ordering is not perfect, there is a risk that >>>you will have to search many moves before you get the cutoff. When you >>>use ETC, you check the hash values for *all* child nodes before you >>>start searching, which can sometimes save a lot of nodes. >>> >>>To me, ETC has always been a clear win. The last time I made any >>>experiments, it reduced my tree size by about 10% at high search depths. >>>I am fairly sure it is a technique which works better with MTD(f) than >>>with more conventional search algorithms, though. >>> >>>Tord >> >>Well, not really. We have different programs but you're right I forgot to >>mention the move ordering part. >>I don't sort moves at all. I generate them in stages and play them before the >>next stage is generated. The stages are hashmove, good captures and so on... >>In the last stage (if no cutoff was produced before) I generate all of the >>remaining moves and pick the 3-5 (depending on prev. history) "best" moves in >>order but without sorting. After that I just take the moves in the order they >>occur in the list. >>This is what I've tested with ETC: >> - In the stage picking the 3-5 moves I go through the moves >> and probe the hashtab using the hashkey they would produce. >> - If a child FH is found handle it and take the next move. >> - If a child FL is found the current node is finished. >> - If the probe announce a hash hit inside the A/B window, >> assign that value to the move and act accordingly >> (in some cases finish the current node) >> - Otherwise go on as before >> >>I haven't seen any win at all even if the tree shrinks. On the contrary it is a >>loser for me... >> >>/Peter > >I have yet to test genuine ETC and if Tord tested that it worked (a clear >winner), I have reasons to GUESS it has a good chance to work ... FOR SOME >TYPE OF IMPLEMENTAIONS. Pseudo ETC already shows great promise for me and paying >good dividends simply because probing etc... before actually making a move is >cheap in my case. > >I think it MAY NOT pay dividends in your program. >I recently made some changes to my move-ordering which I considered >important and correct and the impact seems significant. My raw suspision is >hashmove, good catures, etc is NOT ENOUGH for ETC; it may need to be >complemented by a CLEARLY SELF-SUFFICIENTLY GOOD MOVE ORERING before it works. >If I get you correct, your approach shortcut on move-ordering instead of doing >more move-ordering analysis with attack tables, etc. that Tord may be doing, an >entirely different approach that may not be compatible with ETC. > >Another thing. I'm not sure we need to test all moves. I happen to be >doing lazy eval() before makemove so this itself prunes likely fail-low >moves and if coupled with a good analysed-type move-order that pushes likely >good moves front, it may have a good return... speculation at the moment. > >Rasjid Of courcse you should try it. My and Tord's approaches are completely different and the different experiences of ETC are not surprising at all. I think my ordering is not bad but of course it could be better. The reason to stop after 3-5 moves (plus the maybe 5 moves that was made before) is that if no FH appeared until then the whole node is probably a FL node and the order of moves is not important or at least much less important. /Peter
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.