Evolutionary Robotics and Open-Ended Design Automation - Biomimetics: Biologically Inspired Technologies - page 152

Though these principles are independent, they often appear in tandem and hence the confusion:
we tend to speak of useful modules being reused as building blocks, and indeed recurrence of a
pattern may be an indication of its functional modularity, though not a proof of it.
An inherent tradeoff exists between modularity and regularity through the notion of coupling .
Modularity by definition reduces coupling, as it involves the localization of function. Regularity,
however, increases coupling as it reduces information content. For example, if a module is reused in
two different contexts, then the information content of the system has reduced (the module needs to
be described only once and then repeated), but any change to the module will have an effect on both
places. Software engineers are well aware of this tradeoff. As a function is encapsulated and called
from an increasing number of different contexts in a program, so does modifying it become
increasingly difficult because it is entangled in so many different functions.
The tradeoff between modularity and reuse is also observed in engineering as the tradeoff
between modularity and optimality. Modularity often comes at the expense of optimal perform-
ance. Systems that are less modular, that is more integrated, can be more efficient in their
performance as information, energy, and materials can be passed directly within the system, at
the expense of increased coupling. Software engineers are familiar with ''long jumps'' and ''global
variables'' that have this effect, similarly, mechanical products will often achieve optimality of
performance or cost through integration of parts into monolithic components wherever possible.
The increased performance gained by reduction of modularity is often justified in the short term,
whereas increased modularity is often justified in longer time scales where adaptation becomes a
dominant consideration.
It is not clear whether modularity, regularity, and hierarchy are properties of the system being
evolved (i.e., the ''solution''), or of the target fitness specification (i.e., the ''problem''). It may well
be that there is a duality between these viewpoints. The evolutionary computation literature
contains several instances of test functions that are themselves modular (separable, e.g., Royal
Roads [Mitchell, 1996]), hierarchical (e.g., Hierarchical-IFF [Watson and Pollack, 1999]), and
regular (e.g., one-max). It is not surprising then to see corresponding algorithms that are able to
exploit these properties and find the solutions to these problems.
Engineers often go to great lengths to describe design goals in a way that is solution-neutral,
that is it describes target functionality while placing the least constraints on the solution. Indeed
engineering design is notorious for having multiple — even many — solutions to any given
problem, without any solution being clearly superior. The fact that modular, regular, and hierarch-
ical solutions are more attractive is because — we conjecture — the design process itself tends
to prefer those for reasons of scalability. It is therefore plausible that in search of scalable
algorithms for synthesizing solutions bottom up, we should avoid test functions that have an
inherent modular or hierarchical reward, and have these solution properties emerge from the search
process itself.

4.7.2 Research Methodology

Though robotic systems provide an intuitive and appealing substrate to explore many of these
questions, they also pose many difficulties. They are computationally expensive to simulate and
difficult to construct physically. More importantly, like biology, they contain many beautifully
complex but arbitrary details that obscure the universal principles that we are looking for. There
is always a temptation to increase the fidelity of the simulators, adding more biologically
realistic details, in hopes that this would lead to more life-like behaviors. However, it is sometimes
more fruitful to investigate these questions in a simpler, more transparent substrate. In fact we look
for the minimal substrate that still exhibits the effects we are investigating. Many insights can be
gained by looking at these simplified systems, and the lessons learned brought to bear on the
complex problems of practical importance. Thus much of the research in evolutionary design and
evolutionary robotics is disguised as experiments in much more abstract systems.