Author: Robert Hyatt
Date: 09:16:00 02/16/03
Go up one level in this thread
On February 16, 2003 at 06:52:30, martin fierz wrote: >On February 15, 2003 at 22:23:00, Robert Hyatt wrote: > >>On February 15, 2003 at 08:29:54, martin fierz wrote: >> >>>i was thinking of implementing position learning in my checkers program, and was >>>trying to find out how this is typically done. if i understand it properly, >>>here's what people do: >>> >>>1) if the search returns a value much lower than the previous search (define >>>"much lower" as you like), write this position with it's value to the learn-file >>>on disk. >>>2) before every search, stuff positions from the learn file in the hashtable. >>> >>>this all sounds sensible to me, and i can see that this helps. however, i have a >>>stupid question about it. let's imagine that in the initial position A your >>>program is worse and decides to sac some material for a last-chance attack. a >>>few moves later, in position B, it may think that it's compensation is not >>>sufficient and drop it's score and learn this position. another 2 moves later it >>>suddenly sees that it has a perpetual check, and that the move in position A was >>>not to blame at all. i think this scenario is not quite unrealistic, and it >>>seems to me that position learning doesnt work here, because you only learn that >>>position B was bad (when in fact it wasn't). even if you were writing ALL >>>positions to disk after searching them (something i would like to do in >>>checkers), then you still haven't solved the problem: in your next search at >>>position A you will have a hash hit at position B telling you that the correct >>>move is bad, because you can't see further than your hash hit. >>> >>>hmmm. is there any workaround for this? >>> >>>cheers >>> martin >> >> >>This is a well-known issue, and I should add that position learning doesn't >>claim to solve everything. It is really a defense against repeating the same >>game multiple times which humans have been known to do. But as you notice, it >>will not solve every kind of position, because of local maxima issues. IE >>there are some holes you can not climb out because to do so means you first >>have to go for a _lower_ score, to reach the valley floor and start to climb >>up the other side... > >thanks for the answer. i was afraid i was missing something, but that doesn't >seem to be the case then ;-) > >cheers > martin Tony Scherzer and Dave Slate wrote an ICCA paper on this. (That is where I heard of the idea, in fact.) I believe that they discussed the particular problem you mentioned...
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.