dfols

dfols copied to clipboard

Currently, if nsamples>1, objfun is called multiple times in a loop. Allow for extended objfun evaluation objfun(x, nsamples) where the objfun calculation returns all nsamples results in one go (instead...

lindonroberts

Generalized least squares

2

Allow option for generalized least squares (i.e. instead of ||r(x)||^2, have ||r(x)||_A^2, where A is the inverse of the correlation matrix; see https://en.wikipedia.org/wiki/Generalized_least_squares). This captures correlation/heteroskedacity in the residuals.

lindonroberts

Manual termination

1

Implement a custom exception which allows the manual termination of the solver within an objective evaluation. i.e. If evaluating objfun(x) raises a "TerminationException", then just terminate the solver immediately (returning...

lindonroberts

Implement the "escape local minima" feature from Py-BOBYQA [1] into DFO-LS. [1] Coralia Cartis, Lindon Roberts and Oliver Sheridan-Methven, Escaping local minima with derivative-free methods: a numerical investigation, Optimization (2021)....

lindonroberts

When scaling bounds to [0,1], don't set delta_{max} to 1e10, use something smaller

lindonroberts

Regularization/generic loss function

2

Modify DFO-LS to allow different loss functions (not just sum-of-squares), when the analytic form is known, so full model can be built using first derivatives (e.g. currently, have y ->...

lindonroberts

Starting information

1

Allow DFO-LS to use any prior information, such as: pre-existing evaluations of the objective and/or estimate of Jacobian at x0.

lindonroberts

MATLAB interface

1

DFO-LS should have a MATLAB interface which mimics fminunc/fminsearch. MATLAB's Python interface doesn't support callable functions (e.g. objfun), so need to go through the C interface to both languages. The...

lindonroberts