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

MboPlotDistToNeighbor

Description
The plot shows the distance between the search space components We use Gower distance to handle discrete search space components as well. $$ d_{ij}^{Gower}=\frac{\sum_{k=1}^pw_k\delta_{ij}^{(k)} d_{ij}^{(k)}}{\sum_{k=1}^pw_k \delta_{ij}^{(k)}} $$ where:
$w_k$   weight of the $k$-th search space components
$\delta_{ij}^{(k)}$  0 when variable $k$ is missing or NA, else 1
$d_{ij}^{(k)}$  is the distance contribution of the $k$-th variable to the total distance, i.e. x[i,k] to x[j, k]
$i,j = 1, 2\ldots, n$  observations for each search space component, i.e. rows of design matrix
$p$       number of search space components x

reference: taken from documentation of function [daisy()] (https://www.rdocumentation.org/packages/cluster/versions/2.1.0/topics/daisy) (pkg. cluster)

• $dist_measure \in {min, max, mean}$: Once the distances are calculated we get a lower triangular matrix $M$ with $n$ (number of iterations) rows and $p$ (number of search space components) columns. The $dist_measure$ is applied to each row of $M$, i.e. giving the $min,max, mean$ of each row which is then plotted.

Interpretation
* In each iteration the distance between the values of search space components is calculated (i.e. the plot shows at iteration i how close the search space components $\mathbf{x}^{(i)}$are to each other). * Comparing 2 adjacent points: How far are the proposed search space components (sample points) away from the sampled search space components in the previous iteration.
* Choosing the initial design: If there is no significant drop after the initial design, one may choose a bigger initial design.

How to use
* Trade-off exploration/exploitation: If distance shows 'peak' this might indicate that the optimizer explores leading to a combination of values of the search space components which is far from the previous values.
* Adjustment of infill criterion (if infill criterion bears parameter measuring the degree of exploration): If heavy peaks also at later iterations one may consider to adjust the infill criterion towards a lower degree of exploration to yield better results (as the optimizer then exploits more often).

Modifications * $dist_measure$: We can use different $dist_measure$s from the dropdown in the sidebar panel.
* $include_init_desgin$: We can exclude the points of the initial design by setting $include_init_desgin=FALSE$.

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