Go to navigation (press enter key)Menu
Cognitive Science
Completed Project

Does Neural Network Modularity Enhance Evolvability in Robots?

Cognitive Science and Biology
Jonathan Snyder ’18, David Wallach ’17, and Xiaoqing Xu ’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

Evolvability, the capacity of a population to adapt to environmental change, is of interest not only to evolutionary biologists but also to computer and cognitive scientists who apply functional principles from the evolutionary process to create intelligent software and autonomous systems. What makes one population more evolvable than another? One hypothesis is that structural modularity (Qs) — the degree to which the connections within a network are clustered — enhances evolvability by facilitating sub-networks with specialized functions. Computer simulations have shown that higher Qs is correlated with higher evolvability in artificial neural networks (ANN). However, the causal relation between Qs and evolvability is not clear. Our simulations of ANNs functioning as robotic neurocontrollers, which connect sensors to motors, show that for one given value of Qs there are multiple possible patterns of connections and each pattern functions differently. Thus we hypothesize that some other feature of the ANNs, rather than Qs, is responsible for differences in evolvability. We test this hypothesis using Vassar’s Tadro robotic system, a behaviorally autonomous surface swimming agent, in an environment with spatially distinct areas representing food, shelter, and a danger zone, each marked with a different type of environmental cue. We will evolve two populations of Tadros identical initially in their distribution of Qs values for their ANNs. After ten generations of selection for behaviors that enhance food gathering and shelter finding we will drastically alter the environment by adding the danger zone. After this first round of adaptive evolution, we predict that the more evolvable population, measured by the rate adaptation in the new environment, will have ANNs able to combine information from all the sensors, rather than higher values of Qs.