A new paper by Kashtan et al. in PLoS Computational Biology presents an interesting study of the evolution of modularity, extending their previous work showing "that modular structure can spontaneously emerge if goals (environments) change over time, such that each new goal shares the same set of sub-problems with previous goals."
The evolution of modularity is a topic of longstanding interest in GP and evolutionary computation more generally, within which we often seek to evolve modular programs or structures. Many also seek to leverage the modularity of representations to accelerate evolution. A lot of the work on automatically defined functions, etc., has been concerned with these issues and I think that cross-fertilization with the new computational biology results could be fruitful.
The closest thing that I know of in the GP literature to the Kashtan et al. results is a paper by Terry Van Belle and David Ackley in GECCO 2002, in which they observed the "evolution of evolvability in experiments using genetic programming to solve a symbolic regression problem that varies in a partially unpredictable manner." Alan Robinson and I were inspired by this to do a similar experiment in PushGP, which allows modularity to arise from scratch via code self-manipulation, and we wrote it up briefly in a GECCO 2002 Workshop paper (see section 3.2).
No comments:
Post a Comment