# Difference between revisions of "Changelog"

From SUMOwiki

(→6.2 - October 2009) |
(→6.2 - 6 October 2009) |
||

Line 2: | Line 2: | ||

== 6.2 - 6 October 2009 == |
== 6.2 - 6 October 2009 == |
||

− | |||

− | '''This list is not complete and subject to change at any time, without prior notice''' |
||

* A new neural network modelbuilder "ann". This is a lot faster than the existing "anngenetic" and the quality of the models is roughly the same |
* A new neural network modelbuilder "ann". This is a lot faster than the existing "anngenetic" and the quality of the models is roughly the same |

## Revision as of 16:49, 10 October 2009

Below you will find the detailed list of changes in every new release. For a more high level overview see the Whats new page.

## Contents

- 1 6.2 - 6 October 2009
- 2 6.1.1 - 17 April 2009
- 3 6.1 - 16 February 2009
- 4 6.0.1 - Released 23 August 2008
- 5 6.0 - Released 6 August 2008
- 6 5.0 - Released 8 April 2008
- 7 4.2 - Released 18 October 2007
- 8 4.1 - Released 27 July 2007
- 9 4.0 - Released 22 June 2007
- 10 3.3 - Released 2 May 2007
- 11 3.2 - Released 9 Mar 2007
- 12 3.1 - Released 28 Feb 2007
- 13 3.0 - Released 14 Feb 2007
- 14 2.0 - Released 15 Nov 2006

## 6.2 - 6 October 2009

- A new neural network modelbuilder "ann". This is a lot faster than the existing "anngenetic" and the quality of the models is roughly the same
- The sample selection infrastructure is now much more powerful, sample selection criteria can be combined with much more flexibility. This opens the way to dynamic variation of sampling criteria.
- Support for Input constraints / multiple output sampling in the LOLA-Voronoi sample selection algorithm
- Support for auto-sampled inputs (e.g., frequency in an EM context) in LOLA-Voronoi. This is useful if a particular input is already sampled by your simulator.
- Automatic filtering of samples close to each other in CombinedSampleSelector
- Support for TriScatteredInterp in InterpolationModel when it is available (Matlab version 2009a and later)
- Sample selectors that support it (for example: LOLA-Voronoi) now give priorities to new samples, to that samples are submitted and evaluated in order of importance.
- Support for pre-calculated Latin Hypercube Designs, these will be automatically downloaded and used where possible and will improve performance
- The Blind Kriging models have been improved and can now also be used as ordinary Kriging models. Since these models are superior to the existing DACE Toolbox models, the DACE Toolbox backend has been removed.
- The EGOModelBuilder (do model parameter optimization using the EGO algorithm) now uses a nested blind kriging model instead of one based on the DACE Toolbox. This allows for better accuracy
- The Kriging correlation functions can now be chosen automatically (instead of only the correlation parameters)
- Support for multiobjective optimization in the EGO framewok (extended version of probability of improvement)
- DelaunaySampleSelector, OptimizeCriterion support the same set of criterions
- EGO Improvement criteria can now be used together with DACEModel, RBFModel, and SVMModel (LS-SVM backend only)
- Added a model type and builder that does linear/cubic/nearest neighbour interpolation
- All error functions and measures now consistently deal with complex valued data and multiple output models
- Various improvements in the Model Info GUI as part of the Model browser tool
- Improved stability in LRMMeasure, a behavioral complexity metric to help ensure parsimonious models
- The profiler GUI has been updated and improved, and support for textual profilers has been added.
- Improved performance when using Measures, especially for models with multiple outputs.
- Improved management of the best model trace, also in pareto mode
- Removed the debug output when using (LS-)SVM models and added compiled mex files for Windows
- Ported the remaining classes to Matlabs Classdef format
- Increased use of the parallel computing toolbox (if available) in order to speed up modeling
- Improved the Matlab file headers so the help text is more informative (always includes at least the signature)
- Support for plotting the model prediction uncertainty in the model browser (only for 1D plots and not supported by all model types)
- Added support for so-called "reference by id" on every level of the config. If a tag of a particular type is defined on top-level with an id, it can be referenced everywhere else, instead of copying it entirely. See rationalPoleSupression sample selector and patternsearch Optimizer, for example.
- EmptyModelBuilder added - in case you just want to use the sequential design facilities of the toolbox, but not its models.
- Various cleanups and bugfixes

## 6.1.1 - 17 April 2009

- Various cleanups and bugfixes (see Known bugs for 6.1)

## 6.1 - 16 February 2009

- The default error function is now the Bayesian Error Estimation Quotient (BEEQ)
- Full support for multi-objective model generation, multiple measures can now be enforced simultaneously. This can also be applied to generating models with multiple outputs (combineOutputs = true). Together with the automatic model type selection algorithm (heterogenetic) this allows the automatic selection of the best model type per output.
- The model browser GUI now supports QQ plots
- The Gradient Sample Selection Algorithm has been renamed to the Local Linear Sample Selector (LOLASampleSelector)
- The modelbuilders have been refactored and some removed. This is a result of the optimizer hierarchy being cleaned up. Adding a new model parameter optimization routine should now be more straightforward.
- The interface classes have been renamed to factories as this is more correct. All implementations have been ported to Matlab's new Classdef format and the inherritance hierarchy has been cleaned up. It should now be significantly easier to add support for new approximation types.
- The ModelInterfaces are now known as ModelFactories, this is more correct. Note that the XML tagnames have been changed as well.
- The Model class hierarchy has been converted to the new Classdef format. This means that models generated with previous versions of the toolbox will no longer be loadable in this version.
- The heterogenetic model builder for automatic model type selection has been cleaned up and made more robust.
- Rational models now support all available modelbuilders. This means that order selection can be done by PSO DIRECT, Simulated Annealing, ... instead of just GA and Sequential.
- New optimizers added are (they can also be used as model builders): Differential Evolution
- Added a Blind Kriging model type implementation as a backend of KrigingModel
- Addition of an EGO model builder. This allows optimization of the model parameters using the well known Efficient Global Optimization (EGO) algorithm. In essence this uses a nested Kriging Model to predict which parameters should be used to build the next model.
- Trivial dependencies on the Statistics Toolbox have been removed
- Added a new smoothness measure (LRMMeasure) that helps to ensure smooth models and reduce erratic bumps. It works best when combined with other Measures (such as SampleError for ANN models)
- Models now have a simple evaluateDerivative() method that allows one to easily get gradient information. The base class implementation is very simple but works. Models can override this method to get more efficient implementations.
- Added experimental support for the Matlab Parallel Computing Toolbox (local scheduler only). This means that when the parallelMode option in ContextConfig is switched on, model construction will make use of all available cores/cpu's.
- Many speed improvements, some quite significant.
- Various cleanups and bugfixes

## 6.0.1 - Released 23 August 2008

- Fixed a number of (minor) bugs in the 6.0 release

## 6.0 - Released 6 August 2008

- Many important bugs have been fixed that could have resulted in sub-optimal models
- Addition of a Model Browser GUI, this allows you to easily 'walk' through multi-dimensional models
- Moved the InitialDesign tag outside of the SUMO tag
- Some speed improvements
- Removed support for dummy inputs
- Measure scores and input/output names are saved inside the models, allowing for more usable plots
- Added the project directory concept, each example is now self contained in its own directory
- #simulatorname# can now be used in the run name, it will get replaced by the real simulator name
- Input dimensions can be ignored during sampling if the simulator samples them for you. This is useful in EM applications for example where frequency points can be cheap.
- Logging framework revamped, logs can now be saved on a per run basis
- The global score calculation has changed! it is a weighted sum of all individual measures. (the weights are configurable but default to 1)
- Added a simple polynomial model where the orders can be chosen manually
- Countless cleanups, minor bugfixes and feature enhancements

## 5.0 - Released 8 April 2008

- In April 2008, the first public release of the
**Surrogate Modeling (SUMO) Toolbox**(v5.0) occurred. - A major new release with countless fixes, improvements, new sampling and modeling algorithms, and much more.

List of changes:

- Fixed the 'Known bugs' for v4.2 (see Wiki)
- data points now have priorities (assigned by the sample selectors)
- Vastly reworked and improved the sample evaluator framework
- robust handling of failed or 'lost' data points
- pluggable input queue infrastructure to make advanced scheduling policies possible

- The number of samples to select each iteration is now selected dynamically, based on the time needed for modeling, the length of one simulation, the number of compute nodes available, ... A user specified upper bound can till be specified of course.
- Model plots are now in the original space instead of the normalized ([-1 1]) space
- The default error function is now the root relative square error (= a global relative error)
- Intelligent seeding of each new model parameter optimization iteration. This means the model parameter space is searched much more efficiently and completely
- Added a fast Neural Network Modeler based on FANN (http://fann.sf.net)
- Added a Neural Network Modeler based on NNSYSID (http://www.iau.dtu.dk/research/control/nnsysid.html)
- The LS-SVM model type has been merged with the SVM model type. The SVM model now supports three backends: libSVM, SVMlight, and lssvm
- Added a SampleSelector using infill sampling criterions (ISC).
- The expected improvement from EGO/superEGO is provided among others. (only usable with Kriging and RBF)

- More robust handling of SSH sessions when running simulators on a remote cluster
- The TestSamples measure has been renamed to ValidationSet
- The Polynomial model type has been renamed to the more apt Rational model
- The grid and voronoi sample selectors have been renamed to Error and Density respectively
- Drastically reduced memory usage when performing many runs with multiple datasets (datasets are cached)
- Added utility functions for easily summarizing profiler data from a large number of runs
- Lots of speed improvements in the gradient sample selector
- The default settings have been harmonized and much improved
- The (LS)SVM parameter space is now searched in log10 instead of ln space
- Added a TestMinimum measure
- compares the minimum of the surrogate model against a predefined value (for instance a known minimum)

- Added a MinimumProfiler
- tracks the minimum of the surrogate model versus the number of iterations

- Movie creation now works on all supported platforms
- Added an optimizer class hierarchy for solving subproblems transparantly
- Cleaned up the structure of all the model classes so they no longer contain an interface object. This was confusing and led to error prone code. Virtually all subsref and subassgn implementations have also been removed.
- The MinMax measure is now enabled by default
- The Optimization framework was removed (and replaced) for various reasons, see: http://sumowiki.intec.ugent.be/index.php/FAQ#What_about_surrogate_driven_optimization.3F
- Fixed the file output of the profiler, formatting is correct now
- New implementation of a maximin latin hypercube design
- Minimizes pairwise correlation
- Minimizes intersite distance

- Removed dependency of factorial design on the statistics toolbox
- Added a plotOptions tag, this allows for more customisability of model plots (grey scale, light effects, ...)
- Profiler plots can now also be saved as JPG, PNG, EPS, PDF, PS and SVG
- Countless cleanups, minor bugfixes and feature enhancements

## 4.2 - Released 18 October 2007

- Fixed the 'Known bugs' for v4.1 (see Wiki)
- Simulators can be passed options through an <Options> tag
- Added a fixed model builder so you can manually force which model parameters to use
- Removed ProActive dependency for the SGE distributed backend
- Improved Makefile under unix/linux
- Data produced by simulators no longer needs to be pre-scaled to [-1 1], this can be done automatically from the simulator configuration file
- Deprecated the optimization framework. It is currently under re-design and a better, more integrated version, will be released with the next toolbox version.
- Lots of cleanups, minor bugfixes and small feature enhancements
- In October 2007, the development of the M3-Toolbox was discontinued.

## 4.1 - Released 27 July 2007

- Fixed the 'Known bugs' for v4.0 (see Wiki)
- Vastly improved test sample distribution if a test set is created on the fly
- Gradient sample selector now works with complex outputs and has improved neighbourhood selection
- Speed and usability improvements in the profiler framework
- Improvements in the profiler DockedView widget (added a right click context menu)
- Addition of some new examples
- Added an option (on by default) that selects a certain percentage of the grid sample selector's points randomly, making the algorithm more robust
- Some cleanups, minor bugfixes and feature enhancements

## 4.0 - Released 22 June 2007

- IMPORTANT: the best model score is now 0 instead of 1, this is more intuitive
- Reworked and improved the model scoring mechanism, now based on a pareto analysis. This makes it possible to combine multpile measures in a sensible way.
- Added a proof of concept surrogate driven optimization framework. Note this is an initial implementation which works, but don't expect state of the art results.
- Cleanup and refactoring of the profiler framework
- The profiling of model parameters has been totally reworked and this can now easily be tracked in a nice GUI widget
- Cleanup of error function logic so you can now easily use different error functions (relative, RMS, ...) in the measures
- Improved model plotting
- Support for the SVMlight library (you must download it yourself in order to use it)
- Added a MinMax measure which can be used to suppress spikes in rational models
- Support for extinction prevention in the heterogenetic modeler
- Fixed warnings (and in some cases errors) when loading models from disk
- Respect the maximum running time more accurately
- Many cleanups, minor bugfixes and feature enhancements

## 3.3 - Released 2 May 2007

- Fixed incorrect summary at the end of a run
- Fixed bug due to duplicate sample points
- Ability to evaluate multiple samples in parallel locally (support for dual/multi-core machines)
- Speedups when reading in datasets
- Added 2 new modelbuilders that optimize the parameters using;
- Pattern Search (requires the Matlab direct search toolbox)
- Simulated Annealing (requires Matlab v7.4 and the direct search toolbox)
- The Matlab Optimization Toolbox (includes different gradient based methods like BGFS)

- A new density based sample selction algorithm (VoronoiSampleSelector)
- New simulator examples to test with
- Addition of a profiler to generate levelplots
- Ability to generate Matlab API documentation using m2html
- New neural network training algorithms based on Differential Evolution and Particle Swarm Optimization
- It is now possible to call the toolbox with specific samples/values directly, e.g., go('myConfigFile.xml',xValues,yValues);
- Many minor bugfixes and feature enhancements

## 3.2 - Released 9 Mar 2007

- Many important bugfixes
- Documentation improvements
- Fully working support for RBF models
- New measure profilers that track the errors on measures
- Many new predefined functions and datasets to test with. We now have over 50 examples!

## 3.1 - Released 28 Feb 2007

- Small bugfixes and usability improvements
- Improved documentation
- Working implementation of a heterogenous evolutionary modelbuilder
- More examples

## 3.0 - Released 14 Feb 2007

- Availability of pre-built binaries
- Extensive refactoring and code cleanups
- Many bugfixes and usability improvements
- Resilience against simulator crashes
- Ability to set the maximum running time for one sample evaluation
- Vastly improved Genetic model builder + a neural network implementation
- Addition of a RandomModelBuilder to use as a baseline benchmark
- Possible to add dummy input variables or to model only a subset of the available inputs while clamping others
- Improved multiple output support
- outputs can be modeled in parallel
- each output can be configured separately (eg. per output: model type, accuracy requirements (measure), sample selection algorithm, complex handling flag, etc)
- mutliple outputs can be combined into one model if the model type supports this

- Noisy (gaussian, outliers, ...) versions of a given output can be automatically added
- New and improved directory structure for output data
- New model types:
- Kriging (based on the DACE MATLAB Kriging Toolbox by Lophaven, Nielsen and Sondergaard)
- Splines (based on the MATLAB Splines Toolbox, only for 1D and 2D)

- Now matlab scripts can be used as datasources (simulators) as well
- New initial experimental design
- Based on a dataset
- Combination of existing designs
- Based on the complexity of different 1D fits

- Addition of new datasets and predefined functions as modeling examples

## 2.0 - Released 15 Nov 2006

- Initial release of the M3-Toolbox - open source