The arboretum procedure

Default:

NORANDSCALE, which sets each scale estimate to the standard deviation

of the corresponding target variable.

Note:   NORANDSCALE overrides whatever is specified in the

RANOPTIONS statement.  

Copyright 2000 by SAS Institute Inc., Cary, NC, USA. All rights reserved.

---

The NEURAL Procedure

INPUT Statement

The INPUT statement allows you to group together input variables having common levels and

standardizations. You can specify as many INPUT statements as you want up to the limits imposed

by computer memory, time, and disk space. The input layers can be connected to hidden or output

layers using CONNECT statements.

Category Action Statement - affects the network or the data sets. Options set in an action statement

affect only that statement.

INPUT variable-list / ID=name

<LEVEL=value>

<STD=method>;

Required Arguments

variable-list

Specifies the input variables.

ID=name

Specifies the identifier for the layer.

Options

LEVEL=value

Specifies the measurement level, where value can be:

NOMINAL|NOM

Nominal

ORDINAL|ORD

Ordinal

INTERVAL|INT

Interval

Default:

Interval (for variables specified by a VAR statement in the DMDB

procedure) or nominal (for variables specified by a CLASS statement in the

DMDB procedure).

STD=method

Specifies the standardization method, where method is:

---

NONE|NO

Variables are not altered.

STD

Variables are rescaled to have a mean of 0 and a standard deviation of 1.

RANGE|RAN

Variables are rescaled to have a minimum of 0 and a range of 1. This standardization is not

recommended for input variables.

MIDRANGE|MID

Variables are rescaled to have a midrange of 0 and a half-range of 1 (that is, a minimum of

-1 and a maximum of 1).

Default:

STD

Copyright 2000 by SAS Institute Inc., Cary, NC, USA. All rights reserved.

---

The NEURAL Procedure

NETOPTIONS Statement

Identifies the network options to set.

Category Option Statement - does not directly affect the network, but sets options for use in

subsequent action statements. The options persist until reset at a later stage in the

processing.

Alias: NETOPTS |NETOPT

NETOPTIONS network-option(s);

Network Options

DECAY=number

Specifies the weight decay.

Range:

number   0

Default:

For the QPROP optimization technique: .0001; for all others: 0

INVALIDTARGET=action

Specifies the action taken during training if an out-of-range target value is found, where action

can be:

OMITCASE | OMIT If INVALIDTARGET = OMITCASE is specified, and an invalid target

value is found in the training data set, a warning is given, the observation

is not used, but the training will continue.

STOP

If INVALIDTARGET = STOP is specified, an error is issued, and

training is terminated.

Example:

If ERROR = GAMMA is specified in a target statement, the target values

should be positive. If a zero or negative value is found for the variable(s)

listed in the target statement, the observation containing that value cannot be

used for training. The training either continue with the remaining valid

observations or stops depending on the INVALIDTARGET= specification.

Note, however, if the {\it mean}, over the training data set, of the target

variable(s) is zero or negative, an error is issued and the training is stopped,

regardless of the INVALIDTARGET= specification.

---

Default:

The default is INVALIDTARGET = STOP. INVALIDETARGET can be

abbreviated as INVALIDTARG or INVTARG.

OBJECT=objective-function

Specifies the objective function where objective-function can be one of the following:

DEV

Requests deviance (for ERROR=NORMAL, this is least squares).

LIKE

Requests negative log-likelihood.

MEST

Requests M estimation.

Default:

Depends on the error functions that the network uses. To determine the

default, examine the table below named Errors by Objective Functions.

Scan the table from left to right. The default is the first column that

contains a "yes" in every row corresponding to an error function used

in the network. If no such column exists in the table, an error message

is issued and the network cannot be trained.

Errors by Objective Functions

ERRORS

DEV

LIKE

MEST

Normal

Yes

Yes

No

Cauchy

No

Yes

No

Logistic

No

Yes

No

Huber

No

No

Yes

Biweight

No

No

Yes

Wave

No

No

Yes

Gamma

Yes

Yes

No

Poisson

Yes

Yes

No

Bernoulli

Yes

Yes

No

Binomial

Yes

Yes

No

Entropy

Yes

Yes

No

---