Iteration Controls


Description Usage See Also

Description

Some Draco regression procedures are performed using iterative techniques.  These regressions offer the ability to specify convergence criteria for the iterative procedure.  The user-specified criteria both allow for optimum solution time as well as protection from infinite looping.

Usage

In the Options Menu of regressions using iterative procedures, there will be an Iterations and Convergence... item.  Selecting this option will present the user with the following dialog:

Iteration Control Dialog

Three options are presented by default to the user.  The first, Maximum Iterations, specifies the maximum numer of solution iterations that will be performed regardless of convergence.  If converence to the absolute tolerancedoes not occur during the iterations, the procedure will be stopped when the maximum number of iterations is reached.  This maximum protects against runaway computations, but allows the user to set the limit if convergence is proceeding slowly.

The second, Absolute Tolerance, specifies the convergence criteria at which the regression is considered sucessful.  The tolerance in Draco is compred to the sum of the absolute values of the change in all estimates.  When this sum falls below the absolute tolerance, the regression is considered converged and successful.  

The third, Relaxation Factor, adjusts the multiplier used when applying incremental changes to estimates prior to performing a subsequent iteration.  A relaxation factor of 1.0 means that the computed deltas being applied to estimates is applied as-is.  Often it may be advatageous for slowly converging estimates to set the relaxation factor higher (over-relaxation) to speed up convergence; factors of 1.1 and 1.2 can make significant differences in time to convergence.  On the other hand if a system appears unstable, setting a relaxation factor less than one can dampen unstable behavior so that convergence can be achieved.  

See Also

Two Stage Least Squares
Least Squares (Autocorrelated)
Nonlinear Least Squares
Logit Regression


Copyright © 2008 Approximatrix, LLC
Text licensed under the Creative Commons Attribution-Share Alike 3.0 License
DracoTM and the Approximatrix logo are trademarks of Approximatrix, LLC
Other trademarks are property of their respective owners