June 12, 2012 18:12 PSP Book - 9in x 6in 07-Junichi-Takeno-c07
Theory of Robot Evolution 117
For now, the four surviving robots reproduce eight robots by
mating. These eight robots are the (n + 1)th generation. The
robots of this new generation are evaluated to select the surviving
robots. At this time, the value of the fitness function increased
to more than 77. We can hopefully expect the fitness function to
increase gradually by repeating the above process generation after
generation. Repetition stops when robots of a high evaluation are
found. This is a breakthrough approach in that the details of the
artificial neural network are omitted and the overall picture is
grasped. This sounds like saying that good results are important no
matter how the artificial neural network is structured. That is why
this approach is enthusiastically hailed by behaviorists.
The machine evolution approach is said to have been initiated
by John Holland and L.J. Fogel, both in the United States, and Ingo
Rechenberg in Germany, in the 1960s almost simultaneously (Pfeifer
and Sceider, 2001). Holland advocated a genetic algorithm, while
Fogel promoted evolutionary programming, and Rechenberg de-
veloped an evolutionary strategy technique. The three researchers
seem to have been working in different fields of study, but in reality
there is no significant difference among them from the broader
perspective of machine evolution.
The machine evolution approach has made a great progress
historically hand in hand with the study of artificial life.
“Artificial life,” as defined by Christopher Langton of the Santa
Fe Institute, “is an artificial system that behaves like a living
organism.” In the study of artificial life, artificial living organisms,
called creatures, reproduce themselves, grow, prosper, decay, and die
by simulation in computers.
In their studies, artificial life researchers naturally discuss the
evolution of their artificial living organisms. However, their research
is performed in an ideal environment of computers, which is
markedly different from the environment of machine evolution
research.
Researchers of machine evolution claim that machines acquire
knowledge through interaction with the natural environment. This
very belief prevents them from attempting to build a natural
environment on a computer, which remains a decisive drawback.