What's Changed
FitBenchmarking v0.2.0 is the result of 10 months of work, and we've introduced a lot of changes to the FitBenchmarking package.
This release now requires Python 3.7.1 or greater.
New Features
- Support for different cost functions was added; see below
- Improved support for different jacobians; see below
- Better handling of conflicting options
- More descriptions of the methods in docs
- Box constraints supported
- Licences of the software used documented
- Minimizers are collected into types
- Second derivatives (Hessians) supported, and can be compared
- Added an option to select algorithms by type
- Results directory can be set from the command line
- A maximum runtime can be selected for each solver
Parsers
- Mantid's Crystal Field objective is supported
- Native parser has been refactored
- IVP parser has been added
Fitting Software
- `iminuit>2.0` is supported
- `levmar` supported
- `Matlab` supported (basic, and `curve_fitting`, `statistics` and `optimization` toolboxes)
- `scipy_go` supported
- `gradient_free_optimizers` supported
Cost functions
- `NLLS` and `WeightedNLLSs` cost functions replace the defaults
- `HellingerNLLS` cost function
- `Poisson` cost function added
- Multiple cost functions can be compared at once
Jacobians
- `numdifftools` Jacobians have been added
- Analytic Jacobians added for NIST and SIF file formats
- Solver default Jacobians have been added
- Multiple Jacobians can be compared at once
Tables
- the generation of these has been refactored
- Failed fits are highlighted in the tables
- Customizable colourmaps in the tables
- Header row and leading column of results table is frozen
- Problem descriptions displayed in fitting report
- Accuracy and runtime displayed in fitting report
- Problem summary page has been added
Data sets
- more SASView examples added
- Data Assimilation (IVP) examples added
- CrystalField examples added
- Global optimization SIF file examples added
Breaking Changes
- Support is dropped for `iminuit<2.0`
- `lm-scipy-no-jac` option under `scipy_ls` in `[MINIMIZERS]` has been removed (this is now available under the solver default jacobian)
- `results_dir` has moved from `[PLOTTING]` into `[OUTPUT]` in the options file
- `use_errors` has been removed as an option, and this functionality is controlled by the choice of cost function
- The Jacobian method `SciPyFD` is no longer supported; this has been renamed to `scipy`
- `colour_scale` option in `[PLOTTING]` is no longer supported; this is replaced by the `colour_map`