Author: Bruce Moreland
Date: 09:18:58 02/10/99
Go up one level in this thread
On February 10, 1999 at 09:04:02, Robert Hyatt wrote: >On February 09, 1999 at 13:41:10, Jay Scott wrote: > >> >>On February 07, 1999 at 18:30:17, Robert Hyatt wrote: >> >>>... when I do a probe, I _know_ I will get a result. So >>>the search will definitely terminate here, with win/lose/draw scores. But I >>>can't really go on until I know which... because of alpha/beta.... >> >>In principle you could go on nondeterministically, the same way modern >>processors go on through a conditional branch before they've calculated >>whether the branch is taken. >> >>For tablebase lookups, the drill would be: >> >>- start the lookup; it'll take a while >>- predict the value you'll get back >>- keep on searching as if the predicted value were correct > >this is the point. there is _no_ more searching here, period. Because if I >probe, I am going to hit. And if I hit I won't search further. But I can't >easily exit this node yet because I don't have the score. And with a >depth-first search, trying to 'roll back' is essentially impossible, because >the tree space is discarded as it is searched. > >And I don't see a good way to 'predict' the outcome anyway. The most common >'value' is DRAW. but predicting that would be bad... > > > > >>- when the lookup finishes: >> - if the prediction was good, you're set >> - if the prediction was wrong, you have to roll back and start over: >> - reset the search to that node >> - zap any bad hash entries you made along the way >> - undo any other changes... > > >all of that would likely require a reverse tree search, which means waiting >2x as long before you can use the table value. > > > >> >>I think this would be incredibly bug-prone. I wouldn't dare try it! >>But if you can get the code correct, it would definitely be faster. >> >> Jay > >I'm not sure it would be 'faster' as "unrolling" a depth-first search means >searching in the reverse direction, something i can't even do in the current >version of crafty... This is an imaginative idea, but I doubt it would save much, since you either do so few probes that your NPS is still very high, or so many that you aren't compute bound. And, as was noted, the code would be incredibly intricate. bruce
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.