This page contains a description of the latest updates of RHEIA.
15. What’s new
15.1. v2.1.0.
This release extends uncertainty quantification, improves optimization post-processing and makes sample generation for external models easier to use.
15.1.1. Uncertainty quantification
Lognormal stochastic inputs are now supported in
stochastic_space.csvwithLognormalas distribution type.Lognormal samples written to
samples.csvand passed to the model remain physical lognormal values. Internally, RHEIA transforms them to the corresponding latent Gaussian variables for Hermite PCE construction.Existing
samples.csvfiles with lognormal physical samples are transformed in the same way before fitting the PCE.Model evaluations are appended to
samples.csvone by one and flushed immediately, so completed evaluations are preserved if a run stops early.UQ now validates that the number of outputs returned by the model matches the number of entries in
'objective names'.A
NO_MODELcreate-only workflow example was added to generate PCE training samples without a connected model.
15.1.2. Optimization and post-processing
PostProcessOpt.get_hypervolume()calculates the hypervolume of the Pareto front for every generation.Hypervolume post-processing supports mixed minimization and maximization objectives through objective weights.
The deterministic optimization tutorial now includes a hypervolume convergence plot.
15.1.3. Testing
Added workflow coverage for create-only PCE training samples with
NO_MODEL.Added an end-to-end lognormal equivalence test comparing a lognormal input model against an equivalent latent-Gaussian model.
15.2. v2.0.0.
This release modernizes the package and input formats and adds sparse polynomial chaos expansion support.
15.2.1. Uncertainty quantification
Sparse PCE is now available through the uncertainty quantification dictionary with
'uq method': 'sparse'.Sparse PCE training size is configured with
'n samples'.Full PCE remains the default through
'uq method': 'full'.UQ result files now distinguish full and sparse PCE results. Sparse files include the number of training samples in the filename, e.g.
sparse_pce_order_2_lcoh_n_samples_80.txt.The post-processing class
PostProcessUQacceptsmethodandsamplesto read either full or sparse PCE results.The PCE implementation uses SciPy’s maintained quasi-Monte Carlo tools for Sobol’ sequences.
15.2.2. Input files
design_space.csvnow has a required header:name,type,value,upper_bound.stochastic_space.csvnow has a required header:name,relation,distribution,deviation.Parameter rows in
design_space.csvkeepupper_boundempty; variable rows require bothvalueandupper_bound.All packaged case files and tutorial design-space files have been migrated to the headered schema.
Reading and writing of design-space and stochastic-space CSV files now uses pandas and validates required columns explicitly.
15.2.3. Optimization and post-processing
Latin Hypercube Sampling now uses SciPy’s maintained sampler.
Optimization post-processing uses pandas for sorted CSV output and validates generation consistency in
fitness.csvandpopulation.csv.UQ post-processing uses pandas for PDF, CDF and Sobol’ CSV files and reports clearer errors for malformed PCE summaries.
15.2.4. Packaging
Package metadata and dependencies moved to
pyproject.toml.Legacy
setup.pyandmanifest.inwere removed.RHEIA now requires Python 3.11 or newer.
The documentation build installs the local checkout instead of the published PyPI package.
A
.gitignorewas added for common Python, build, IDE and generated result files.
15.3. v1.1.11.
This was the previous stable version of RHEIA.
15.4. v1.1.5.
The input files for the cases (design_space.csv and stochastic_space.csv)
and the output files for optimization (fitness.csv, population.csv)
and uncertainty quantification (samples.csv, full_pce_order_x_y_Sobol_indices.csv) are converted from extensionless files into CSV files.
15.5. v1.0.0.
This is the official release of RHEIA! It corresponds to the package published in the Journal of Open Source Software.