How to Select Sampler
Guidelines for Selecting a Optimization Algorithm Settings
Much of the content is sourced from the official documentation of Optuna, the optimization library used by Tunny.
Flow Chart
This may not always be the best choice, but in general, the following flow can be used to select which sampler to use.
Sampler info table
Advanced Usage
Optimization using sampling results
Bayesian optimization TPE and GP require constant sampling with Random or Quasi-MonteCarlo before optimization due to the characteristics of the optimization method.
By default, a random sampler is automatically used to sample the set "Number of startup trials".
However, if you first select any sampler and sample a certain number of samples, then select TPE or GP and perform the optimization again, you can use that sample data to perform the optimization.
Reveal the importance of variables
Tunny is not just an optimization tool, but a design support tool. For example, you can identify the importance of variables and see which variables have the greatest impact on the design.
Below are the results of the Hyper Parameter Importance visualization introduced in the Visualization Tab. This shows the effect of each variable on the objective function 0, which sums to 1 for each variable.
For this type of study, use Quasi-MonteCarlo instead of TPE, GP, NSGAII, or CMA-ES to perform the optimization, as it is easier to obtain accurate values up to higher dimensions.