A block-coordinate Gauss-Newton method for nonlinear least squares

Add to your list(s) Download to your calendar using vCal

• Jaroslav Fowkes (University of Oxford)
• Thursday 22 November 2018, 12:00-13:00
• Nuffield G13.

If you have a question about this talk, please contact Sergey Sergeev.

Nonlinear (nonconvex) least squares problems are used for a range of important scientific applications, such as data assimilation for weather forecasting and climate modelling, where parameter estimation is needed in order to specify simulation models that fit observations. In many of these applications, a model run is computationally expensive but provides the full vector of simulated observations. However, calculating the entire Jacobian may be too expensive, as it may involve additional model runs along each variable coordinate. We propose a block-coordinate Gauss-Newton method that calculates Jacobians only on a subset of the variables/parameters at a time. We investigate globalising this approach using either a regularization term or a trust-region model and show global complexity results as well as extensive computational results on CUT Est test problems. Furthermore, as our approach exhibits very slow rates of convergence on certain problems, we design adaptive block size variants of our methods that can overcome these inefficiencies.

This talk is part of the Optimisation and Numerical Analysis Seminars series.

This talk is included in these lists:

• Nuffield G13
• Optimisation and Numerical Analysis Seminars
• School of Mathematics Events

Note that ex-directory lists are not shown.