The Mountain Car Problem - Design of Experiments for Reinforcement Learning - page 102

6

11

13

25

32

ʱ mag

ʳ

ʱ ratio

ʻ ʵ

FAST99 Eff Jansen

Fig. 5.6 First-order sensitivity indices for the mountain car domain. The subregion number is shown
on the left side of the plots.

In addition to the first-order sensitivity indices, we use functional ANOVA
(FANOVA) graphs to assess two-way interaction effects between the parameters.
Figure 5.7 shows the FANOVA decomposition, where each node represents one pa-
rameter, the size of each node is scaled to the first-order sensitivity index, and the
width of the line between nodes is scaled to the size of the interaction between pairs
of parameters. It should be noted that the size of the nodes and the width of the
interaction lines should not be compared across subregions, as these are scaled to the