"Dener, Alp via petsc-users" <[email protected]> writes:
> About Levenberg-Marquardt: a user started the branch to eventually contribute > an LM solver, but I have not heard any updates on it since end of April. For > least-squares type problems, you can try using the regularized Gauss-Newton > solver (-tao_type brgn). The problem definition interface is a bit different. > BRGN requires the problem to be defined as a residual and its Jacobian, and > it will assemble the gradient and the Hessian on its own and feed it into the > standard Newton line search solver underneath. There are also a few different > regularization options available, like proximal point Tikhonov (l2prox) and a > sparsity term (l1dict) that can be set with the > (-tao_brgn_regularization_type) option. These get automatically cooked into > the gradient and Hessian. Combined with the automatic diagonal shifts for the > Hessian, BRGN really is almost equivalent to an LM solver so it might do what > you need. That'd be here, based off your previous comments. https://gitlab.com/tkonolige/petsc/-/commits/brgn_lm Tristan graduated last month so I don't know if he'll be doing further legwork on this. I'll see if he can make a placeholder MR.
