ABSTRACT
This article discusses the requirements for evolvability in complex systems, using random Boolean networks as a canonical example. The conditions for crystalization of orderly behavior in such networks are specified. Most critical is the emergence of a "frozen component" of the binary variables, in which some variables are frozen in the active or inactive state. Such frozen components across a Boolean network leave behind functionally isolated islands which are not frozen. Adaptive evolution or learning in such networks via near mutant variants depends upon the structure of the corresponding "fitness landscape." Such landscapes may be smooth and single peaked, or highly rugged. Networks with frozen components tend to adapt on smoother landscapes than those with no frozen component. In coevolving systems, fitness landscapes themselves deform due to coupling between coevolving partners. Conditions for optimal coevolution may include tuning of landscape structure for the emergence of frozen components among the coadapting entities in the system.
