The Evolution of Everything (EvE) and Genetic Programming 143
3 The Evolution of Everything (EvE)
What if we combined the locality and real-world data flow made available by
the Internet of Things with the evolutionary power of genetic programming? How
would we design such a system? What would be the result?
Since GP is usually used with a limited set of inputs and a very focused purpose
as represented in a fitness function, what happens when the inputs are expanded to
include a significant part of the real world and the purpose is broadened to a more
general and less single-valued focus? The IoT makes it possible to conceive of such
a situation and to expand the application of GP to more open-ended problems.
Paradoxically, it is the added constraints of the real world that increase the
usefulness of GP. These include the constraints of physics, resource allocations and
timeliness which may be difficult to capture in an artificial environment but are
intrinsically present in a real-world setting.
3.1 Non-trivial Geography
In Spector and Klein ( 2005 ), the authors describe a GP system where they use a very
simple notion of locality only for the purpose of crossover. While many GP systems
use an amorphous breeding population where any individual can combine with any
other, Spector and Klein added a simplistic, linear structure to the geography of
the population to constrain the potential breeding population to a limited number of
individuals within a finite distance from an individual as potential breeding partners.
This simple change produced solutions that were much more effective in outcome
(i.e., better fitness) and more economical (i.e., less computational effort to produce)
than results produced without a notion of geography.
While the notion of geography was not new to evolutionary algorithms since,
as Spector and Klein noted, others had used complex simulation environments with
distinct geographies or sub-populations (demes) that were complex in representation
and difficult to manage, Spector and Klein showed that the use of even a simple
geography seemed to lead to significant improvement at minimal cost. In essence,
they had developed a reductio ad absurdum test to the notion of individual
locality and the result had been a startling increase in performance. Their tentative
conclusion from a GP perspective was that limiting the mating pool increased
diversity among locales. They also point out that because of the simplicity of
implementation, it could easily be adopted in systems that already have a notion of
geography (such as deme based systems) or even in systems where sub-populations
are defined by performance, such as ALPS (Hornby 2006) or other performance
segregated sub-populations.
In the IoT, many devices have a distinct location in the real world. For example,
farms, water sources, power sources and cars all attach great significance to location.
By tying GP selection for crossover to real-world locations, GP would find localized