In PhilippScheller/VisBayesOpt: Visualizing model based optimization runs from package mlrMBO

MboPlotInputSpace

Description
For each search space component, the plot shows the histograms over the evaluated values. The top (bottom) shows numeric (discrete) search space components.

Interpretation
* Prior: Values (artificially) sampled according to the specified input space (considering eventual transformations of the input space). This artificially shows the regions that would have been searched over by a random search.
* Posterior: Values actually evaluated by the optimizer.

How to use
* Comparing prior and posterior: Enables to see how the optimizer searches compared to random search. Shows which range of a single search space component the optimizer focuses on.
* Analyzing posterior: If the optimizer searches heavily on an upper/lower boundary of a specified range one may adjust the range of the search space components as it seems the optimizer finds the minimum $f(x)$ at a different range of $x$.

Modifications * $include_init_design_sampling_distribution$: We can exclude the sampled observations from the specified prior of the search space components by setting $include_init_design_sampling_distribution = FALSE$ at the sidebar panel.

PhilippScheller/VisBayesOpt documentation built on Sept. 14, 2020, 12:47 p.m.