Manual Page for Command >>> LEVENBERG
Syntax
>>> LEVENBERG: lambda
Parent Command
>> CONVERGE
Subcommand

Description
This command sets the initial value of the Levenberg parameter lambda
(default: 0.001). During the optimization process (see command
>>> LEVENBERGMARQUARDT>) the Levenberg parameter is divided by the Marquardt
parameter nue (see command >>> MARQUARDT) after each successful iteration,
and is multiplied by nue if the new parameter set leads to an increased value
of the objective function, i.e., if an unsuccessful step was proposed.
A big value for lambda means that a small step along the steepest descent
direction is performed. A lambdavalue of zero is equivalent to a GaussNewton
step. The former is robust but inefficient, the latter has a quadratic
convergence rate but may lead to unsuccessful steps.
Example
> COMPUTATION
>> CONVERGE
>>> maximum number of ITERATIONS: 10
>>> set initial LEVENBERG parameter to: 0.1 to make a
safe first step
<<<
<<
See Also
>>> MARQUARDT
Back to Command Index
Page updated: July 29, 1997