Uses of Neural Networks in the SAS System

Size: px

Start display at page:

Download "Uses of Neural Networks in the SAS System"

Asher McKinney
5 years ago
Views:

1 ,:'". ~.,~,~.. ~'"-:!'<~:k~t,,?_,sll\l:5!4::~.. t.j_tili5ej!i[!qn'fl-ft..;..{t'i.,.."l!,.~!'i"":'n~~""",,,.,.d5~_"j!d'i!p~::::!!'t... ~~'1~~~->1'~"'-~'J;':..;T"'-:::,...,,...,.::~.i " ~"~' ~ :VM',~-j~~--;,...:::,::t.:-'::'-:: ~--=<.~;-.!'~'i"'!,..--:-" -.,-g.f,.-'::-- 1~ 1"\ I'l '~ Abstract: Uses of Neural Networks in the SAS System By Martin Duffy & Derek Powell In the past 5 years Neural Networks have moved from the world of academia ~ to being a recognized data analysis tool for business. Neural Networks can be N applied to a wide range of business analysis requirements, but in the past neural analysis tools have NOT addressed a diversity of industries. This paper intends to show how Neural Networks in the SAS System can address a very broad range of business needs, such as CPE, direct marketing and credit. scoring.

2 ~1~~'\~~K-~~~Hi~~}~~~~...,'>{_'i~~ ~~~~':':~:-,~~'.' ~'f.ry~~'.,l~l<'~7~...-;~i'i'\-\"':':""'1~~':':.,;:?7."7~,... I'-~":.,~-,~?,."':"_'"i-~-_~:-.:-~-;~;:~~~t...';~w'''..:.-.. _-,:-!-~~"""-.'":_''-._,~'t',.+.,.,..~,.,_..'~>o _. Introduction Lets start with a brief overview of Neural Networks Welcome to the Bluffer's guide to Neural Networks. This guide will try to address the questions that most people have about neural networks but are too afraid to ask. ~ What is a neural network? (Choose one) a) b) c) d) What your brain is made of A method of spotting patterns in data A method of predicting future outcomes based on historical information. All of the above

3 '::., " rt~;o,;f.~-:).m'l<.~"i,,"~«('~.!'~:'i'!i.'$;;:;h!\irisl.o:;,.... tj,..."" "'<A:.. i'~~.~~~"'~'~ I7-\~~~~ '~."i.l~fj.~~""1"","''''';'i.p.''''''':- cr.;-::;...~~~~...,;,.", ~r, -,.,,:,--,:,-.... ~"""", ",,. '<" '.'"",,,,..,. ' ~ -'~ ",V~"'f.~~'. I!~ I,' I~i l,. How Simple are Neural Networks? What is a neural network? (Answer) The correct answer is of course (d). Neural networks take historical data and "learn" about patterns in it. They can then use the patterns ~ that they have learned to predict what will happen in the future, or to classify data into one of many categories. You do this already, for example if it is snowing outside you predict it is cold. What the neural network does is try to mimjc,t.. he way your brain learns. The only problem is that your brain h~s billions of times the number of neurons the largest neural network has, even if you have been drinking German Beer.

4 ~~7~m~~~~:!t.f\':::~;.'t:","';'_~:;'f"'"f:-{ii;~~1~~.~~~_~':f~-:<.~~:"--{~~~'t;;;~~_t-{~?":'~."J,!,<"'f~}-~~':'f"_".3~"'t":'j:~.<.,,:_-::~"'---:-;:~~~.. ~.~l'~:.:'!-'?"" "'7,~>-~ -:r";"'.t-"'::~"k"',,\~~~ < :.' _""."'<_. _-".",...,. -L --~~. - -" ~-..., What is a Neural Network? Artificial Neural Networks are a ~iae class of flexible models including: ~ Flexible nonlinear regression and Discriminant models Data reduction models Nonlinear dynamical systems A Neural Network is best defined as a set of simple, highly interconnected processing elements or neurons that are capable of "Learning" information presented to them. ~-"~~-:'.'-

$~_< _-,;o<#';"''''i'_~_~~'' :\''i,,,,,,_,,,_,,_~~rt~~'''~''''''~t~1''\.,,!:,~''''="-''''''::-'''''}f-_"~ 'E'j>:~W1!'\i"'<,'~'L~~_-_~_~~~~,-'_q 5'_''-O:;::~-~_''':-~~:-,'--''_"" ;~_,,!$

5 ~_< _-,;o<#';"''''i'_~_~~'' :\''i,,,,,,_,,,_,,_~~rt~~'''~''''''~t~1''\.,,!:,~''''="-''''''::-'''''}f-_"~ 'E'j>:~W1!'\i"'<,'~'L~~_-_~_~~~~,-'_q 5'_''-O:;::~-~_''':-~~:-,'--''_"" ;~_,,!;3C_:_'_-::::-_~'1_";_-i-'_-'~-'~ --- ~ '-"'- \~' ;:!.;; I~ ) ~l,i What do N eural Nets Look Like? ~ 0\ Structure of a Network A neural network in its basic form is composed of several layers of neurons; an input layer, one or more hidden layers and an output layer. Each layer of neurons receives its input from the previous layer or from the input layer. The output of each neuron feeds the next layer or the output layer of the network. Below is a figure of a neuron followed by a figure of a simple 3 layer neural network showing how the neurons fit together. A Neuron

6 ?<.Li~':l!,~~f~.h~~t-.:"~~~i~:";:t~l-f:.;~f~l1R--~~~~~;"'~;x'~~i;:~~~!-~_'1~~"r'1~~)t~~'<'7i"f~f~?:~:;C~":~'f,~}':>5r.;:~-..:::;:.r.::-, <._.~..._.m)'1""~ :("",F" ;.,-.-~,.:-...,- """~'-:.'";""'-'_...'" ". '_' f~'".~..,.~~l~"'>'"~'_ What do Neural Nets Look Like? Inputs, Hidden layers and Outputs A Three layer Neural Network ~ -...J A simple 2x3xl neural network.,.. -,-:c,-~.~... -,--_. --~ ---., '- --'-~"',-~ ~---~----

7 ~--'_,...,_~... ~'.Olit~~\\~.<FS::._t!!... fb.1fi4ullim.:""':,.. d!l"~~~~~~~.l'-~~:=o'~-'~7n"7,:.""'!""".n:""i~j<: ~~~"".""(:""<"".. "", -T, -,"'...,~~_-.~, :~-:.-""'- -;;--" t~ " -")... - r ~ ' ~"'."-' '..,.,,-" ~,... ~-~----"..~-.- tl :~ Tuning the Radio IY()~+?ti.t1 y _ 21~ """-----'" Picture if you can a Neural Net working like a radio receiver and the type of network you choose corresponds to the type of radio waveband we will select. Therefore we use different types of NN' s to work on different types or volumes of data. ~ Type of Net or Waveband Back Prop. Radial Basis An 1\ " 1\ " 1\ On V\l VvllVV 4 A.M. F.M. l;-rr""»-"--

8 ~~f<~~~~~~~#ffl~~i.~"!~~tr~:1fl1':~.!hn~~~~~m~,~:-'~'-:,~!~;,k:~:... ~~,,:-~~h';''-'.',,~'>-:;~oe:r.t~.-tf!''~tf~3?i'::~~~:.!\~~~~'':''"... ~~-'::~~ :''"'''''t:.rc:~~.h":;'~i~~~.;'l!'}-,.':'c,... rl:,:~",:,-:,;c~"'a..-;-,r~ :-,-",-,,,,,. "'c, Tuning the Radio (Continued) Initial Random Seed or Tuning The initial random seed can be viewed as your first attempt to tune into the information in the data. The initial random seed is a start point in a range of random values that will be given a starting ~ values to the synaptic connections (or weights) in the network. 10 ~ Synaptic Connections p

9 ~_~'''''''';>( ~'',.''ol:.,; ",,~~~...'... ~~~::t''!-~~-'-'!?(~f~"-~~~':''''-l-.,.,';:_-,.')":""f.}.":}.:;-:... ;.,;'v';~_'::""i.>,-::-~.~... :~...,...;,~~.;~.".'-;~:t.~7"-i.,,~_"c_~:;-::':/"~i'"'.,.'~,'~~o':'1""7~.,.,."~,'."'... ~M"~:<.. ",.'~.'... '.-.-~,.'. Tuning the Radio (Continued) Neurons in Hidden Layer or Radio Frequency Changing the number of neurons in the hidden layer can be viewed as tuning into a strong frequency within your data or tuning into a trend or pattern within the data. 00 -o Number of Neurons in Hidden Layer = 3

10 ~~~~i.~~'*..~~~~~~~7~~~.~s:~~~~k:t}':.~~:-j'(;,:,~~;n'~~"e<'::9~~'"'-,,,\t! Z ~~~~?O:~""-;':F~,:--;-c:~~...,...,-"'.!:~~",.~ :;, "":,-.. '.'":'=;n:'";"c -~;:-T;"'~f.""'i-~ ~ -:~"">.-'''.~ Tuning the Radio (Continued) Transfer Function or Mono/Stereo 00 - There are many types of transfer functions that can be associated with ne\lrons, two typical transfer functions are Sigmoid and Linear. These transfer functions affect the speed and nature of the networks ability to 'fit' the data. Sigmoid Transfer Function Linear Transfer Function "'\.. f(x)=lll * exp(-x) f(x)=x Stereo Mono

11 ;-.:._...,', '~.,~.~r!:c~'!~~~%"nitlzs'bffiq}i"',.rr~~i"~~~~i."'h :";);-~"-f..-::""',"'~2':'''' -:''''<;f'<-;;.-:.~':'~',~~?,,: ~~~~'l:fi''';':; r.j.-:.,.,...:;:.:t-;..:-:::-;r... ~.-.c-.:-.-.;-~"""'t1:;;j.t:~-_"'f.>'5.'~.~, ,i"~.. "."'.'~~;:'.. _-..;:~-_ "-",~'... ~ ,--. 1l :a Tuning the Radio (Continued) Convergence Value or Fine Tuning The convergence value is how close we want the neural net to fit the data, so this can be viewed as a parallel to fine tuning the radio to a particular frequency for the strongest signal. ()O -N. A Small Convergence Valne i.e may find a strong signal A Large ConvergenceValne i.e may find a weak signal,\// ;~:.. ---"'.

MINE 432 Industrial Automation and Robotics

MINE 432 Industrial Automation and Robotics Part 3, Lecture 5 Overview of Artificial Neural Networks A. Farzanegan (Visiting Associate Professor) Fall 2014 Norman B. Keevil Institute of Mining Engineering