Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!godot.cc.duq.edu!news.duke.edu!news.mathworks.com!newshost.marcam.com!zip.eecs.umich.edu!umn.edu!tera.cs.umn.edu!dedmonds
From: dedmonds@tera.cs.umn.edu (Dirk W. Edmonds)
Subject: CMAC 
Message-ID: <D4w2uB.CwK@news.cis.umn.edu>
Sender: news@news.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: tera.cs.umn.edu
Organization: University of Minnesota, Twin Cities
Date: Sat, 4 Mar 1995 00:02:59 GMT
Lines: 22

Hi netters,

In Albus' 1975 paper "Data Storage in the CMAC" from the Transactions of 
the ASME,  he give an example using a 1 dimensional function (a sine wave),
an input space of size 360, and a neighborhood of 32.  Now, it 
seems to me that you only need a matrix of size 392 to store all the association
cells, since there is only one association cell difference between neighboring 
inputs.  It seems hashing is hardly necessary, since the data can be stored directly in a table of size 392.  But he uses hashing anyways, and the amazing 
thing is, he gets a fair amount of collision, even though his hash table is 
size 4096.  The only way I could get significant collision was with a hash table
of size 392.

Can someone explain?

Dirk

-- 
Dirk Edmonds			University of Minnesota
Graduate Student		Department of Computer Science
Office 612/626-7503		EE/CSCI Building
AIRVL  612/625-8544		200 Union Street SE
dedmonds@cs.umn.edu		Minneapolis, Minnesota 55455
