The arboretum procedure

Mbernoulli

Yes

Yes

No

Multinomial

Yes

Yes

No

Mentropy

Yes

Yes

No

RANDF=number

Specifies the degrees of freedom parameter for random numbers. See the following

Randomization Options and Default Parameters table.

RANDIST=name

Specifies the type of distribution to be used for random initial weights and perturbations. The

distributions and default parameter values are as follows:

Randomization Options and Default Parameters

RANDIST

RANLOC

RANSCALE

DF

NORMAL

mean=0

std=1

-

UNIFORM

mean=0

halfrange=1

-

CAUCHY

median=0

scale=1

-

CHIINV

-

scale=1

df=1

Default:

NORMAL

RANDOM=integer

Specifies the random number seed.

Default:

0

RANLOC=number

Specifies the location parameter for random numbers. See the above Randomization Options and

Default Parameters table.

Specifies the scale parameter for random numbers. See the above Randomization Options and

Default Parameters table.

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

---

The NEURAL Procedure

NLOPTIONS Statement

Identifies the nonlinear optimization 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.

NLOPTIONS <nonlinear-options>;

Nonlinear Options

ABSCONV= number

Specifies an absolute function convergence criterion. ABSCONV= is a function of the

log-likelihood for the intercept-only model.

Default:

The default value is the negative square root of the largest double precision

value.

Range:___number'>Range:

number > 0

ABSFCONV= number

Specifies an absolute function convergence criterion.

Default:

0

Range:

number > 0

ABSGCONV= number

Specifies the absolute gradient convergence criterion.

Default:

1E-5

Range:

number > 0

ABSXCONV= number

Specifies the absolute parameter convergence criterion.

Default:

0

Range:

number > 0

---

DAMPSTEP= number

Specifies that the initial step size value for each line search used by the QUANEW, CONGRA, or

NEWRAP optimization technique cannot be larger than the product of number and the step size

value used in the former iteration.

Default:

2

Range:

number > 0

DIAHES

Forces the optimization algorithm (TRUREG, NEWRAP, or NRRIDG) to take advantage of the

diagonality.

FCONV= number

Specifies a function convergence criterion.

Default:

1E-4

Range:

number > 0

FSIZE= number

Specifies the parameter of the relative function and relative gradient termination criteria.

Default:

Not applicable.

Range:

number   0

GCONV= number

Specifies the relative gradient convergence criterion.

Default:

1E-8

Range:

number > 0

HESCAL= 0 | 1 | 2 |3

Specifies the scaling version of the Hessian or cross-product Jacobian matrix used in NRRIDG,

TRUREG, LEVMAR, NEWRAP, or DBLDOG optimization.

Default:

1 - for LEVMAR minimization technique

0 - for all others

INHESSIAN= number

Specifies how to define the initial estimate of the approximate Hessian for the quasi-Newton

techniques QUANEW and DBLDOG.

---

Default:

The default is to use a Hessian based on the initial weights as the initial

estimate of the approximate Hessian. When r=0, the initial estimate of the

approximate Hessian is computed from the magnitude of the initial gradient.

Range:

number > 0

INSTEP= number

Specifies the initial radius of the trust region used in the TRUREG, DBLDOG, and LEVMAR

algorithms.

Default:

1

Range:

number > 0

LCEPS | LCEPSILON= number

Specifies the range for active constraints.

Range:

number > 0

LCSINGULAR= number

Specifies the tolerance for dependent constraints

Range:

number > 0

LINESEARCH= number

Specifies the line-search method for the CONGRA, QUANEW, and NEWRAP optimization

techniques.

Default:

2

Range:

1   number   8

LSPRECISION= number

Specifies the degree of accuracy that should be obtained by the second and third line-search

algorithms.

---