5 ms·
This is why you actually do something like Levenberg-Marquardt, which combines a gradient descent with a trust region of quadratic optimization, which is varied
by snovv_crash 8y ago
This is why you actually do something like Levenberg-Marquardt, which combines a gradient descent with a trust region of quadratic optimization, which is varied depending on if the cost function is decreasing or not. And of course it is iterative, so a single bad step doesn't destroy everything. The advantage comes from faster convergence in practice.
- mi_lk 8y agoDo you know of any good literature on LM or trust region? I'm interested in the details.
- snovv_crash 8y agoI strongly recommend reading the documentation for Ceres Solver [1]. It is intuitive and in-depth, and references everything you will need. 1. http://ceres-solver.org/nnls_solving.html#trust-region-methods http://ceres-solver.org/nnls_solving.html#trust-region-metho...