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

Modularity and the Evolvability of Artificial Network Control Systems in Embodied Robots

Evan Altiero, Vassar College ’16, John Loree, Vassar College ’16, Jessica Ng, Vassar College ’16, Joshua Ridley, Vassar College ’17 and Profs. Joshua Bongard, Ken Livingston, Nick Livingston, John Long, Jodi Schwarz and Marc Smith

Evolvability is a species’ potential for evolutionary change. One hypothesis suggests evolvability may evolve in concert with modularity. This is based on the argument that if a population has some individuals with highly modular genotype to phenotype (G → P) mappings, then the population will more quickly evolve after exposure to novel fitness landscapes compared to a population without highly modular G → P mappings.  To study the evolution of modularity and evolvability, we use simulated and embodied robots.  We construct simple G→P maps that specify how to build an artificial neural network (ANN) control system that connects the sensors of a robot to its motors.  We use genetic algorithms (GA) to explore how these structures evolve to solve the problem of finding light and then measure the degree of modularity (see poster by Loree, et al.) that evolves under various conditions.

To start the evolution of the population, we create and load randomly generated ANNs into a swimming robot, Tadro14c (see poster by Ridley et al.). Tadros operate autonomously, with sensors that detect light; the structure of the ANN determines how the light detected alters behavior.  In a pool with a single, fixed light source, the amount of light that is gathered is the direct measure of individual fitness. 

The GAs take fitness values as input. The first GA we used was a hill climber (HC), in which the genes from the fittest individuals are bred together to create a new population. Unfortunately the HC is extremely sensitive to local optima. We then used the Age-Fitness Pareto Optimization (AFPO) algorithm, which segments the population of ANNs into different age groups defined by when the solution was generated and evaluates the age groups separately. This allows younger solutions to evolve without direct competition from the more established ones. This segmentation avoids premature convergence on local optima.

Currently our genome encodes for only the ANN. The G→P map is simple (a 1-1 relation). During the coming year we will begin incorporating more complex G→P maps that encode for the Tadro morphology as well as the neural networks.