Information Technology Reference
In-Depth Information
the previous section (compare Tables 7.6 and 7.7 with Table 7.4). Note also
that, for this problem, the presence of random numerical constants results in
a slight increase in performance both in the acellular (average best-of-run
fitness of 5.64991 in the GEA-B experiment against 6.70823 in the GEP-
RNC) and cellular systems (average best-of-run fitness of 4.84324 in the
multicellular system without RNCs compared to 6.01525 in the MCS-RNC
experiment).
Let's now analyze the structures of the best-of-experiment solutions de-
signed in all the four GEP systems in order to compare their simplicity with
the overwhelming complexity of the Kolmogorov-Gabor polynomials cre-
ated in the previous section.
In the GEA-B experiment, the best solution was found in generation 4528
of run 5 (the sub-ETs are linked by addition):
012345678901234
/*+-ch+ghgadbgd
/*+-+h+jigcbieh
-j/-+++deghbbbf
(7.9a)
It has a fitness of 8.29243 and an R-square of 0.9159703184 evaluated over
the training set of 90 fitness cases and, thus, is considerably better than all
the models designed with the STROGANOFF systems. More formally, the
model (7.9a) above can be expressed by the following function:
(
g
h
)
c
(
j
i
)
(
g
c
)
g
h
2
b
y
j
(7.9b)
h
g
a
h
b
i
d
e
In the GEP-RNC system, the best solution was designed in run 62 after
4862 generations. Its genes and respective arrays of random numerical con-
stants are shown below (the sub-ETs are linked by addition):
Gene 0: /*+*-++?cgiiiia68585856
C
0
:
{-0.955719, 0.919098, -0.821198, 0.897492, 0.102966,
-0.827301, 0.504425, -0.117828, -0.410248, -0.597534}
Gene 1: /*++-i+acjibiab56995766
C
1
:
{-0.955719, 0.771973, 0.456207, 0.557037, -0.687439,
0.62616, -0.989319, -0.484497, 0.861939, 0.636414}
Gene 2: *j+j-?jidc?ahbj85025038
C
2
:
{-0.288116, -0.868805, -0.85849, -0.445801, 0.172882,
0.623901, -0.907898, -0.41098, 0.849945, 0.384094} (7.10a)
