Page 130 - 49A Field Guide to Genetic Programming

116 12 Applications

Table 12.1: Samples showing the size and location of Elvis’s finger tip
as apparent to this two eyes, given various right arm actuator set points (4
degrees of freedom). Cf. Figure 12.1. When the data are used for training,
GP is asked to invert the mapping and evolve functions from data collected
by both cameras showing a target location to instructions to give to Elvis’s
four arm motors so that its arm moves to the target.

Arm actuator Left eye Right eye
x y size x y size
-376 -626 1000 -360 44 10 29 -9 12 25
-372 -622 1000 -380 43 7 29 -9 12 29
-377 -627 899 -359 43 9 33 -20 14 26
-385 -635 799 -319 38 16 27 -17 22 30
-393 -643 699 -279 36 24 26 -21 25 20
-401 -651 599 -239 32 32 25 -26 28 18
-409 -659 500 -200 32 35 24 -27 31 19
-417 -667 399 -159 31 41 17 -28 36 13
-425 -675 299 -119 30 45 25 -27 39 8
-433 -683 199 -79 31 47 20 -27 43 9
-441 -691 99 -39 31 49 16 -26 45 13
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
continues for a total of 691 lines

most symbolic regression fitness functions tend to include summing the er-
rors measured for each record in the data set, as we did in Section 4.2.2.
Usually either the absolute difference or the square of the error is used.

The fourth preparatory step typically involves choosing a size for the
population (which is often done initially based on the perceived difficulty of
the problem, and is then refined based on the actual results of preliminary
runs). The user also needs to set the balance between the selection strength
(normally tuned via the tournament size) and the intensity of variation
(which can be varied by modifying the mutation and crossover rates, but
many researchers tend to fix to some standard values).