Does Neural Network Modularity Enhance Evolvability in Robots?
Cognitive Science and Biology
Ben Tidswell ’18, Mackenzie Little ’17, Meghan Willcoxon ’18, Ken Livingston (Cognitive Science), Nick Livingston (Biology), John Long (Cognitive Science & Biology), Jodi Schwarz (Biology), and Collaborators (University of Vermont): Josh Bongard and Anton Bernatskiy
Robots and other artificial agents are often used as models for biological processes such as evolution, learning, adaptation, and natural selection. Of particular theoretical interest to researchers in many disciplines is whether the ability to evolve (evolvability) is itself evolvable. Evolvability is measured as the capacity of a population to vary and survive in the face of environmental changes. Our research tests the hypothesis that the ability to learn increases evolvability using a newly developed genomic coding scheme heavily inspired by biological genetic codes. This hypothesis has primarily been explored in theoretical models or in simulations, while our work uses physically embodied robots. This allows a richer test of the evolution of evolvability by including actual evolutionary change of both body morphology (e.g., number and position of sensors) and the structure of an artificial neural network (ANN). We will test the hypothesis by placing two populations of robots, one with the capacity to learn and one without, in environments complex enough to require behavioral solutions to the exclusive-or problem. This is necessary to evolve sufficiently complex ANNs to engage the evolvability question. The introduction of learning should have the effect of increasing genomic and phenotypic variance compared with a non-learning population, which should in turn increase evolvability.