We are proud to release the SUMO Toolbox 2015a. This new release can be downloaded through the download page and includes many bug fixes and a couple of new features. The SUMO Toolbox 2015a has support for the latest version of Matlab (2015a). We have highlighted a number of important changes and additions below. A more detailed changelog is included at the bottom of this page.
Note: Compatibility fixes for Matlab versions downto 2010a are included. In the future we can not guarantee full compatibility for older Matlab releases. It is simply impossible to test the SUMO toolbox sufficiently across all versions. In addition, handling compatibility issues with the different Matlab versions result in very complex code which is prone to bugs. Fortunately, this only concerns new functionality and the addition of new algorithms, existing code will not be altered and the majority of the toolbox will still be compatible.
As part of this release, support for the nlopt library was added. Several other non-linear optimization methods can now be used, for instance to determine optimal model hyperparameters. The NLOptimizer class is the SUMO wrapper for the library, and has configuration options similar to what is required for NLopt.
While the NLopt library is included with the SUMO toolbox, it is need to be compiled for your platform and Matlab version. Assuming your system has a proper development setup (a compiler, the make utility, Matlab setup) running 'make contrib' from the SUMO toolbox directory will compile NLopt (and other libraries such as libsvm) for your Matlab version. If you do not have the make utility (e.g., windows users) please refer to the NLOpt instructions (in src/matlab/contrib/nlopt/) on how to compile it for your system.
Updated Neighbourhood-Voronoi Sampling Algorithm
The Neighbourhood-Voronoi sampling algorithm has been updated to sample more aggressively near the class boundaries using the 'directedSearch' option (right image). This can be helpful in situations where the focus is on characterising the class boundary as quickly as possible. The behaviour of the algorithm remains as before when 'directedSearch' is turned off (left image).
- Support for Matlab 2015a
- Addition of the NLOPT optimization library. This allows usage of many optimization algorithms (both local/global and gradient based/free algorithms.
- See the new NLOptimizer class
- Addition of a sequential space-filling design (fuzzy-density)
- Added PCTOptimizer: An optimizer which ensures that the ‘best’ candidates (obtained using a user-specified base optimizer) are sufficiently far apart from each other.
- Added HillbertCurve: A space filling curve for use as an initial design where measurements are performed in lines instead of points.
- Removed support for the NNSYSID neural network library
- Removed the MPS algorithm
- Improved the Gaussian Process models
- Optimization of the hyperparameters is much better now
- Updated the GPML toolbox to version 3.4
- Improved the Efficient MultiObjective optimization method
- reference point is chosen more intelligently
- Simplified the random number generator
- Now only a seed is used (see randomSeed option in xml)
- Removal of DACE (deprecated by ooDACE)
- RBF Refactor: FastRBF support removed, code has been updated to fit better within the toolbox
- Addition of a CombiOptimizer to optimize with several optimizers at the same time
- ScatteredDataset also has a native MATLAB implementation
- LOLA-Voronoi algorithm has a 'directedSearch' option to look for classification boundaries more aggressively
- Reworked some ELM code
- Various bugfixes