The Truck Backer-upper Problem - Design of Experiments for Reinforcement Learning - page 116

Table 6.4 Summary statistics for the convergent subregions for the TBU problem.
Subregion # Iteration p conv No. points Dim. ratio Ave. radius Sum radii

12 2 0.917 12 13.415 0.212 2.540

25 3 0.917 12 5.776 0.149 1.792

31 3 0.905 21 12.194 0.160 1.925

32 3 0.939 33 7.947 0.180 2.158

35 3 0.910 78 7.445 0.198 2.370

63 4 1.000 18 13.462 0.109 1.307

70 4 0.967 30 17.778 0.170 2.038

ʱ mag

n hnodes ʱ ratio

Convergent
Non−convergent

Convergent
Non−convergent

Convergent
Non−convergent

KS = 0.186 KS = 0.264 KS = 0.131

11.0 23.5 36.0 48.5 61.0 0.00333 0.00667 0.01000 1.00 2.75 4.50 6.25 8.00

Parameter value Parameter value Parameter value

ʳ

ʵ

ʻ

Convergent
Non−convergent

Convergent
Non−convergent

Convergent
Non−convergent

KS = 0.325 KS = 0.28 KS = 0.316

0.00 0.25 0.50 0.75 1.00 0.900 0.925 0.950 0.975 1.000 0.6 0.7 0.8 0.9 1.0

Parameter value Parameter value Parameter value

Fig. 6.5 Regional sensitivity analysis based on the convergence of all experimental runs for the
TBU problem.

between the cumulative distributions, the more likely that the parameter has an effect
on convergence. Furthermore, the relative shape of the distributions can indicate what
region of the parameter space allows for convergence. For example, the RSA plot
for ʻ shows that the convergent cumulative distribution rises quickly and reaches its
maximum or total value at around ʻ = 0 . 25. This suggests that convergence is more

likely to occur in the lower range of the parameter space. Comparing this qualitative