Levmar

Levmar is an Open Source implementation of the Levenberg-Marquardt optimization algorithm. This method combines the properties of both the steepest decent and Gauss Newton methods to quickly converge upon local solution of a function that is expressed as the sum of squares of non-linear functions.

Levmar may be selected from within the GUI or specified by levmar on the command line. In addition to the standard output files produced at the end of the minimization, this minimizer also generates output on the console that is worth monitoring

Licensing and Acknowledgement

Levmar is distributed under GPL v2 license. The developer (Manolis Lourakis) requests that if you use Levmar in your work you include a citation to the Levmar website. Please see the Levmar website for further details.

Usage

After loading the data and creating a suitable model, the user should select which parameters should be explored (by checking "Free") and then specify the minimum and maximum for each free parameter. All of this may be done in either the GUI or in the JSON save file. Once this step has been completed it is advised that the user save the file and then begin the minimization process.

Because the algorithm finds local solutions, it is suggested you choose your parameter space wisely and inspect the resulting output to ensure Levmar has converted on a valid solution.

Configuration options

Levmar is a very powerful general minimization engine which we have found is particularly sensitive to topology in "chi-squared space." Because of this, we have altered the default configuration for Levmar to be better suited for the models and data sets we have encountered. Unfortunately, we have yet to expose any configuration options to the user in the GUI. If you find that you need to tweak the gradient decent step size or convergence detection criteria, please see the CLevmar::run function in src/minimizers/CLevmar.cpp and the documentation of the parameters on the Levmar website.