Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!gatech!newsjunkie.ans.net!butch!enterprise!news
From: Don Specht
Subject: Origins of GRNN
Message-ID: <1995Jul31.174825.26852@enterprise.rdd.lmsc.lockheed.com>
Keywords: general regression neural networks, GRNN, PNN, Parzen windows
Organization: Lockheed
X-Newsreader: <WinQVT/Net v3.9>
Date: Mon, 31 Jul 95 17:48:25 GMT
Lines: 33

Subject: Origins of GRNN
Prof. Ripley of Oxford and Warren Sarle of SAS have commented in recent postings 
that the General Regression Neural Network (GRNN) is similar to the kernel 
regression of statistical literature.  Prof. Ripley writes, "Kernel discriminant 
analysis has been widely used in medical and chemical applications through the 
work of Hermans, Habbema and co-workers and their ALLOC-80 package.  They have 
been the true enablers of this methodology, about 15 years ago."   
He says that I should have referenced Hermans, et al in my paper, "A General 
Regression Neural Network," which appeared in the IEEE Trans. on Neural 
Networks, Vol. 2, Nov. 1991.  Perhaps I should have to tie things together, but 
not to acknowledge priority.  The GRNN paper is a reformulation of an earlier 
paper in terms of neural networks, with emphasis on the parallel structure which 
can be used in dedicated parallel hardware.  The earlier paper derives what is 
now known as kernel regression plus the Volterra series approximation to it,  
and is:
"A practical technique for estimating general regression surfaces," by D. F. 
Specht, June 1968, Lockheed report LMSC 6-79-68-6, Defense Technical Information 
Center AD-672505, and also available from NASA, aquisition number N68-29513.
Prior to publishing the 1968 paper, we asked famed statistician Prof. Herman 
Chernoff of Stanford Univ. (Dept. of Statistics) to review it for accuracy and 
originality.  He assured us that this was original work and that it was correct. 
 Realizing that the paper needed wider circulation, I submitted it to the 
journal which I thought was the most appropriate in my field of EE, the IEEE 
Trans. on Information Theory in August, 1968.  They rejected the paper, not 
because it was in error, but because "they didn't think that their readers would 
be interested."    The editor was probably right--it was before its time.
Anyone in cyberspace can easily find the original reference by searching NASA's 
recon database
http://techreports.larc.nasa.gov/cgi-bin/NTRS? search_words = "general 
regression".  Clicking on the title will give publication date and ordering 
information from NASA.  Or, if you prefer to drop me a line 
(specht@pc-smtp.rdd.lmsc.lockheed.com), I will send you a copy. 

