ARCtan

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

*The NEURAL Procedure*
**COMBINATION FUNCTIONS**
A combination function combines the values received from preceding nodes into a single number called

the net input. Both output and hidden layers are assigned combination functions.

The following combination functions are available.

**Add**
Adds all the incoming values without using any weights or biases.

**Linear**
Is a linear combination of the incoming values and weights.

**EQSlopes**
Is identical to the Linear combination function, except that the same connection weights are used

for each unit in the layer, although different units have different biases. EQSlopes is mainly used

for ordinal targets.

**EQRadial**
Is a radial basis function with equal heights and widths for all units in the layer.

**EHRadial**
Is a radial basis function with equal heights but unequal widths for all units in the layer.

**EWRadial**
Is a radial basis function with equal widths but unequal heights for all units in the layer.

**EVRadial**
Is a radial basis function with equal volumes for all units in the layer.

**XRadial**
Is a radial basis function with unequal heights and widths for all units in the layer.

The following definitions apply to the Table of Combination Functions:

**All summations**

Are divided by the net inputs indexed by** i**.

The altitude of the **j**th unit

The width (bias) of the **j**th unit

A common bias shared by all units in the layer

The fan-in of the

**j**th unit

The weight connecting the **i**th incoming value to the **j**th unit

The common weight for the **i**th input shared by all units in the layer

The **i**th incoming value

**Combination Functions**
**FUNCTION**
**DEFINITION**
ADD

LINear

EQSlopes

XRAdial

EHRadial

EVRadial

EWRadial

EQRadial

RADial

defaults to EHRadial

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

*The NEURAL Procedure*
**INITIAL Statement**
**After a network has been defined in terms of input, hidden and output layers, all weights and**
**biases in the network must be given initial values before any training is performed. PROC**
**NEURAL will by default supply appropriate random or computed values for these quantities. If**
**you train a network without supplying an INITIAL or USE statement, the network will be**
**initialized using the default specifications.**
**Category **Action Statement - affects the network or the data sets. Options set in an action statement

affect only that statement.

**INITIAL** <

**BIADJUST**=

*adjustment-value*>

<

**INEST**=

*SAS-data-set*>

<

**INFAN**=

*number*>

<

**OUTEST**=

*SAS-data-set*>

<

**RANDBIAS|NORANDBIAS**>

<

**RANDOUT|NORANDOUT**>

<

**RANDOM**=

*integer*>

<

**RANDSCALE | NORANDSCALE**>;

**Options**
**BIADJUST= ***adjustment-value*
Specifies how to adjust the random biases for units with a LINEAR combination function. A

random bias is adjusted by multiplying by the function of the weights indicated by

**adjustment-value**, and dividing by the scale (RANSCALE=) of the distribution from which the

random bias was drawn. **adjustment value** can be one of the following:

SUM

Adjusts random initial biases for the sum of the absolute connection weights leading into

the unit. This value is typically used with STD=MIDRANGE

for inputs and
RANDIST=UNIFORM.

USS

Adjusts random initial biases for the square root of the sum of squared connection weights

leading into the unit. This value is typically used with STD=STD for inputs and

RANDIST=NORMAL.

NONE|NO

No bias adjustment.

**Default:**
BIADJUST=NONE

**INEST=***SAS-data-set*
Specifies an input data set that contains some or all of the weights. Any weights in the INEST=

data set that have missing values are assigned values according to the RANDOM=, RANDOUT,

and RANDBIAS options, as well as the options that pertain to random number distributions that

you specify in the Random statements. An INEST= data set will typically have been created by

using the OUTEST= option in a SAVE or a TRAIN statement from a previous execution of the

NEURAL procedure.

**INFAN=***number*

Divide random connection weights by

(**fan-in ** of unit)** number

where the "fan-in" of a unit is the number of other units feeding into that unit, not counting the

bias or altitude.

**Default:**

0 for radial combination functions, otherwise .5

**Range:**
between 0 and 1

**OUTEST=***SAS-data-set*
Specifies the output data set that contains all the initial weights.

**RANDBIAS | NORANDBIAS**
Specifies whether to randomize output biases.

**Note:** NORANDBIAS overrides whatever you specify in the RANOPTIONS statement.

**Default:**
NORANDBIAS, which sets bias to the inverse activation function of the

target mean.

**RANDOM=***integer*

Specifies the random number seed.

**Default:**
0

**RANDOUT | NORANDOUT**

Specifies whether to randomize the output connection weights.

**Note:** NORANDOUT overrides whatever you specify in the RANOPTIONS statement.

**Default:**

NORANDOUT, which sets weights to 0.

**RANDSCALE | NORANDSCALE**
Specifies whether to randomize target scale estimates.