The Evolution of Everything (EvE) and Genetic Programming 145
Fig. 4 Particulate genes in GP
components tend to be reused across generations means that the results of an
application to a set of data may be cached for reuse during fitness evaluations. It
also avoids some of the intrinsic limits that are associated with tree-based GP as
described in Daida ( 2003 ).
In the context of the IoT, there is an additional extension to the SKGP that
is needed for EvE: the addition of particulate genes. Freeland ( 2003 ) strongly
advocated for an adoption of particulate genes in GP. By ‘particulate genes’ what is
meant in this context is the notion that sections of code and/or values are exchanged
as a whole with a section of code (a particulate gene) taken from one parent or the
other intact. Using the SKGP, particulate genes could be developed where each gene
consisted of a combinator function that could be applied to a specific set of data and,
potentially, be applied to one another. Figure4 shows an example of a particulate
set of genes.
Moreover, one could envision a meta-function applied to a list of particulate
genes in such a way that the meta-function would select and apply some of the
individual gene-expressions to one another based on values produced because of
the context of the values produced in each gene. In this way particulate genes could
mimic the removal of introns during the post-transcription modification of RNA.
Figure6 shows a function being applied to a list of expressions and data in order
to select and combine elements of the particulate gene, effectively combining the
elements in the list according to the function being applied to it.
From an evolutionary perspective, particulate genes have many advantages.
Crossover can now be particulate, with a complete “gene” from either of two parents
being selected and possibly mutated, leaving the structure of each gene intact and
avoiding the disruption of tree-based crossover. It is also easier to flag individuals
as belonging to a particular species, or gender and allows dominant/recessive genes,
all of which improve population dynamics.
3.4 GP Reinforcement Learning (GPRL)
If the SKGP is combined with the ideas behind ERL, while using particulate genes,
then neural networks can be embodied in a single gene and crossover can proceed in
a manner similar to the way it is done in ERL with parent weights being combined
during crossover. Essentially Evaluation neural nets and Action neural nets become
genes in a particulate gene approach to GP.