The NEURAL Procedure
ARCHITECTURE Statement
Constructs a network with 0 or 1 hidden layers, sets the hidden-unit ACT= and COMBINE=
options, and sets default values for various other options as described below.
Interaction: You
cannot override the hidden-unit ACT= and COMBINE= options implied by the
ARCHITECTURE statement, because these are what define the architecture. You can
override all the other values set by ARCHITECTURE by using an INPUT, a HIDDEN,
a TARGET, or a RANOPTIONS statement.
Alias: ARCH
ARCHITECTURE architecture-name
<
HIDDEN=n>
<
DIRECT>;
Required Arguments
architecture-name
Names the architecture you want to use to construct the network. Only one architecture-name
from the following list can be specified:
GLIM
Requests a Generalized Linear Model.
MLP
Requests a Multilayer Perceptron.
ORBFEQ
Requests an Ordinary Radial Basis Function Network with Equal Widths.
ORBFUN Requests an Ordinary Radial Basis Function Network with Unequal Widths.
NRBFEQ
Requests a Normalized Radial Basis Function Network with Equal Widths and
Heights.
NRBFEH
Requests a Normalized Radial Basis Function Network with Equal Heights and
Unequal Widths.
NRBFEW Requests a Normalized Radial Basis Function Network with Unequal Heights and
Equal Widths.
NRBFEV
Requests a Normalized Radial Basis Function Network with Equal Volume.
NRBFUN Requests a Normalized Radial Basis Function Network with Unequal Heights and
Unequal Widths.
Note: See the following two tables for INPUT, TARGET, and HIDDEN options implied by
architecture name and RANOPTIONS implied by architecture name.
Options_Implied_by_Architecture_Name__ARCHITECTURE__NAME__INPUT__Options'>INPUT, TARGET, and HIDDEN Options Implied by Architecture Name
ARCHITECTURE
NAME
INPUT
Options
TARGET
Options
HIDDEN
Options
GLIM
STD=NONE
STD=NONE
No hidden layers
MLP
ACT=TANH
COMBINE=LINEAR
ORBFEQ
ACT=EXP
COMBINE=EQRADIAL
ORBFUN
ACT=EXP
COMBINE=EHRADIAL
NRBFEQ
NOBIAS
ACT=SOFTMAX
COMBINE=EQRADIAL
NRBFEH
NOBIAS
ACT=SOFTMAX
COMBINE=EHRADIAL
NRBFEW
NOBIAS
ACT=SOFTMAX
COMBINE=EWRADIAL
NRBFEV
NOBIAS
ACT=SOFMAX
COMBINE=EVRADIAL
NRBFUN
NOBIAS
ACT=SOFTMAX
COMBINE=XRADIAL
The following definitions apply to the table below:
fan_in
specifies the fan_in of a hidden unit, that is, the number of non-bias and non-altitude
weights feeding into the unit.
n_hidden_units is the number of hidden units.
defloc
is 2*max(.01,(n_hidden_units**(1/fan_in))).
ranloc
is the value of the RANLOC= option.
RANOPTIONS Implied by Architecture Name
ARCHITECTURE
NAME
RANOPTIONS for
BIAS -> HIDDEN
Weights
RANOPTIONS
for INPUT ->
HIDDEN
Weights
(not affected by
early stopping)
RANOPTIONS for
ALTITUDE ->
HIDDEN Weights
GLIM
MLP
ORBFEQ
RANLOC=defloc
RANSCALE=ranloc*.1
RANSCALE=1
ORBFUN
RANLOC=defloc
RANSCALE=ranloc*5
RANSCALE=1
NRBFEQ
RANLOC=defloc
RANSCALE=ranloc*.1
RANSCALE=1
NRBFEH
RANLOC=defloc
RANSCALE=ranloc*.5
RANSCALE=1
NRBFEW
RANLOC=defloc
RANSCALE=ranloc*.1
RANSCALE=1
RANLOC=1
RANSCALE=ranloc*.5
NRBFEV
RANLOC=.5*defloc
RANSCALE=ranloc*.1
RANSCALE=1
NRBFUN
RANLOC=defloc
RANSCALE=ranloc*.5
RANSCALE=1
RANLOC=1
RANSCALE=ranloc*.5
Options
HIDDEN= n
Specifies the number of hidden units for all architectures other than GLIM.
Default:
None
DIRECT
Requests direct connections from inputs to outputs.
Copyright 2000 by SAS Institute Inc., Cary, NC, USA. All rights reserved.
The NEURAL Procedure
CODE Statement
To score a data set based on a previously trained neural network outside of PROC NEURAL, you
can specify a CODE statement to write SAS DATA step code to a file or catalog entry. This code
can then be included into a DATA step that reads (using a SET statement) the data set to be
scored.
Category Action Statement - affects the network or the data sets. Options set in an action statement
affect only that statement.
CODE FILE=
file-name
<
ERROR|NOERROR >
<
FORMAT=
format>
<
GROUP=
name>
<
RESIDUAL|NORESIDUAL>;
Required Arguments
FILE=file-name
Specifies the filename to which the scoring code is written.
Options
ERROR | NOERROR
Specifies whether to generate code to compute the error function.
Default:
NOERROR
FORMAT=format
Specifies the format to use for the weights.
Default:
BEST12
RESIDUAL | NORESIDUAL
Specifies whether to generate code to compute residuals. If you request code for residuals and then
score a data set that does not contain target values, the residuals will have missing values.
Default:
NORESIDUAL
Copyright 2000 by SAS Institute Inc., Cary, NC, USA. All rights reserved.