Ben Kovitz, David Bender, Marcela Poffald
Artificial Life Conference Proceedings
Any part of a genome, considered separately from the rest of the genome, evolves against a “virtual fitness landscape” that results when the rest of the genome is held constant. We show how analyzing a genome in this way can explain one form of progressively increasing evolvability. When one part of a genome is a vector of numbers (“knobs”) and the rest is a graph that determines the mapping from knobs to phenotype, the graph will respond to selective pressure to “acclivate” the virtual fitness function faced by the knobs—that is, to make it more hill-shaped. For as long as the knobs’ virtual fitness function provides opportunity for distorting it to make knob-turning mutations improve fitness, the graph experiences pressure to evolve those distortions as a side-effect of responding to its own virtual fitness function. As the knobs’ virtual fitness function grows more hill-shaped, the knobs track upward paths more easily and hence so does the genotype as a whole. A synergy develops between incremental exploration of phenotypes by knob-mutations and discontinuous exploration by graph-mutations. A favorable condition for this is a global fitness function that frequently varies, changing constants but leaving structural invariants unchanged. The graph then accumulates a memory of the invariants as revealed across many previous epochs, held in the form of bias limiting and directing future evolution.