Evolvability in Developmental Systems

The developmental mapping from genotype to phenotype is responsible for much of the evolvability (adaptive variation) exhibited in nature (Raff 1996; Kirschner & Gerhart 1998; Wagner & Altenberg 1996). Random mutations are transformed into structured phenotypic variation and the effects of deleterious mutations are mitigated. This mechanism is powerful precisely because search becomes constrained, generating only highly adaptive phenotypes. In evolutionary computation, acquiring such constraints and bias is akin to learning the underlying structure of a particular fitness function; such structure can be exploited to improve search efficiency. For example, when evolving a design for a coffee table, an evolvable encoding would discover that table height and surface area correspond to fundamental axes of variation, constraining search to solutions that maintain constant height and high surface area (Hornby 2004). Such evolvability is exhibited in indirect encodings, particularly developmental systems. Evolvability helps find good “trajectories” through the search space, not just good fitness peaks. However, selection for evolvability is not always possible with many fitness functions; in general evolution is opportunistic, and will take large immediate fitness gains over smaller fitness gains that may lead to better (e.g. more evolvable) parts of the search space. Thus selecting for evolvability and selecting for good solutions can be viewed as generally antagonistic goals. In other words, even when using developmental encodings, there may be no real selection pressure to optimize the genomic representation for evolvability. This observation may help explain in general why in some cases developmental encodings do not perform well when compared to simpler direct encodings (Reisinger & Miikkulainen 2006). One way in which a consistent selection pressure for evolvability can be generated is by systematically changing the fitness function over time (Kashtan & Alon 2005). However if the representation itself is not adaptable, as in the case of direct encodings, this selection pressure is ignored, leading to highly-optimal, but “brittle” solutions. We believe that this phenomenon may be responsible for much of the cycling and disengagement behavior seen in competitive coevolution; since direct encodings cannot store information about search, populations may be prone to “forgetting” past