Tunny v0.5.0

Summary

• Constraints can now be taken into consideration.
  • Only TPE, GP, NSGAII can use constraint.
  • Multiobjective optimization can now take constraints into account; visualization of ParetoFront results shows areas that do not satisfy the constraints, as shown in the gray areas below.
  • 
• Add Clusterting output
  • The k-means method can be used to cluster the Pareto solution of the optimization.
  • 
• Add Hypervolume output
  • Hypervolume visualization was added as one of the indicators to determine convergence of multi-objective optimization.
    • If the increase in this value is small, you can consider that the multi-objective optimization has converged.
  • 
• Advanced optimization settings can now be made from the UI.
  • 

Details

Breaking Change

• The format in which optimization results are saved has been changed. Optimization results from previous versions cannot be handled.

Added

• Hypervolume visualization
  • It is useful for determining convergence in multi-objective optimization.
• Clustering visualization
  • Clustering of results during multi-objective optimization makes it easier to evaluate solutions.
• BoTorch Sampler
  • This sampler use Gaussian Process and support multi-objective optimization.
• Quasi-MonteCarlo Sampler
  • Detail about this sampler.
• Support Constraint.
  • Only TPE, GP, NSGAII can use constraint.
• Sampler detail settings UI
  • Previously it was necessary to change the JSON file of the settings, but now it can be changed in the UI
• Enable Text Bake in the FishMarket component.
• Allows selection of NSGA-II crossover methods.
  • Uniform, BLXAlpha, SPX, SBX, VSBX, UNDX
• Ability to set Popsize on CMA-ES restart
• Run GC after trial when has geometry attribute or setting always run.
  • This change probably make optimize slower before
  • If you want to cut this setting, set the value of "GcAfterTrial" to 2 in Settings.json.
• Show LICENSE button in Tunny UI.

Changed

• When genepool is an input, it now creates variable names using nicknames.
• The output of the Pareto solution was made to consider the constraints.
• Multivariate in TPE sampler default option is false to true
• Updated Optuna used to v3.0.0
  • Use suggest_int and suggest_float instead of suggest_uniform for more accurate variable generation in optimization
  • Random and Grid samplers now support multi-objective optimization
  • The format of the db file in which the results are saved has changed. Please note that it is not compatible with the previous one.

Fixed

• The PythonInstaller window now has no text on the progress bar.
• When more than one Study exists, another Study Name is set and RunOpt no longer causes a Solver Error.
• The error does not occur when the Brep of Geometry of Attribute is null.

Related PR

• Feature/more good readme by @hrntsm in #71
• Made the README look cool. by @hrntsm in #72
• Add hypervolume visualization by @hrntsm in #74
• Feature/add botorch sampler by @hrntsm in #76
• Fix/use genepool nickname by @hrntsm in #78
• Feature/support handle integer & discrete_uniform suggestion by @hrntsm in #80
• Fix python installer launch detection by @hrntsm in #83
• Fix/change studyname error by @hrntsm in #84
• Feature/update optuna 300rc by @hrntsm in #85
• Feature/constraints by @hrntsm in #87
• Feature/settings UI by @hrntsm in #91
• Fix pareto front too long attr disp by @hrntsm in #93
• Feature/bake txt in fishmarket by @hrntsm in #95
• Feature/k means by @hrntsm in #100
• Feature/support consider feasible output by @hrntsm in #101
• Feature/update optuna v3 by @hrntsm in #103
• Feature/run-gc-after-trial by @hrntsm in #105
• Add license button by @hrntsm in #106
• Release v0.5.0 by @hrntsm in #107

Full Changelog: v0.4.0...v0.5.0