Prev | Next | whats_new |

Added the routine l1_with_l2 which combines L1 fitting and/or regularization with L2 fitting and/or regularization.

The routine l1_linear was using the operator

`®=`

which is defined in Octave but not in Matlab.
This had been fixed.

The linear_ok.m
test case has been changed to test
scale the problem so it is more difficult and to check the number
of iterations for the solution.

The initial state used by l1_linear
has been improved
so that it is closer to a solution of the first order conditions.
This reduced the number of iterations necessary.
The line search in l1_linear and l1_quadratic was missing a factor of the linear search parameter. This has been fixed.

Revert l1_newton_line to pervious version because it was faster. Then make some further speed improvements.

The convergence criteria was changed to
square root of average squared residual instead of
square root of sum of squared residuals.

The updating of the relaxation factor was changed to occur when the
line search parameter is greater than or equal .9.
In addition, a constant divisor of four was used for the update.
Modified the l1_newton_line function (used by l1_linear ) so that it solves an @(@ m \times n @)@ linear least squares problem instead of inverting a corresponding @(@ n \times n @)@ matrix (numerically more stable).

The documentation was reorganized so that
l1_newton_quad
is below l1_quadratic
and
l1_newton_line
is below l1_linear
.

Remove usages of the `&=`

operator in linear_ok.m
(it is defined in Octave but not in Matlab®).
First version released.

Input File: whats_new.omh