The Evolution of Everything (EvE) and Genetic Programming 147
4 Discussion
The Genetic Programming Theory and Practice workshop was designed to encour-
age speculative ideas. Parts of EvE are highly speculative. Fog Lifter is a work
in progress, but much of the code is drawn from open source software that is quite
robust. The SKGP exists and has been applied to commercial and research problems,
particularly in the biotech space. However, it does not yet use particulate genes,
instead relying on directed graphs to replace trees of standard “Koza Style” GP.
While it is safe to say that these graphs can be contained within the conceptual
notion of a particulate gene, this has not yet been implemented.
Similarly, while ERL as described in Ackley and Littman ( 1991 ) has been
implemented, the author has not had access to the code base. It is probably a good
idea to revisit ERL in this context.
The adaptation that uses particulate GP with specific genes encompassing the
output of neural nets as part of a “post-transcriptional” removal of “exons.” is
(as far as the author has been able to determine) a wholly new idea in the GP
world though there has been some use of the word ‘exon’ to denote “code bloat”
in the GP literature (see for example Soule 2002). Here, the meta-function may, in
one situation, select one set of genes for use in the overall function and in another
circumstance, select others. In this, it must be confessed, the concept is closer to
epigenetic phenomena or splice variants in biology than simply removing exons.
Beyond the mechanics of how EvE might work, there is the underlying idea that
moving from simulated worlds to real-time, real-world data, creates a sea change for
evolutionary algorithms. Instead of being limited by imprecise models of a world
that evolves for a set number of generations, it is based on continuous, adaptive
evolution. While it is not expected that such a system would instantly create a high-
precision model of the real world, over time it is expected that valuable predictive
models will emerge. Moreover, the failures of such systems can teach us much about
the world. The author has long contended that we learn as much or more from
analyzing the failures of GP to make correct predictions as we do from successes
as described in. Assumptions about the real world are often mistaken and GP often
makes these mistakes clear during runs that incorporate these assumptions.
Beyond simply modeling the world, (Holland 1995 , 1998 ), talk at length about
the property of emergence of new properties and behaviors and describes what
he believes is the key components of such systems: State, Transition Function,
Generators and Agents. EvE maps fairly well onto these elements as follows:
- State: the internal state of an individual function in response to the current
environment; - Transition Function: the application of an evolved function in response to the
environment; - Generators: the real world (no need to simulate an environment in the IoT!);
- Agents: individuals in the environment.