Evolutionary Robotics: The Role of Gene Duplication and Modularity in the Emergence of Evolvability
Michael Jaklitsch ’21, Lingxiu Zhang ’21, Jason Han ’23, and Professor Kenneth Livingston (Cognitive Science)
We use simulated robots to test hypotheses about basic principles governing the evolution of intelligent agents. Specifically, we hypothesize that mechanisms of gene duplication and subsequent differentiation increase the modularity of a robotic agent’s sensor/neural network/motor system over generational time, resulting in populations that are more evolvable, that is, better able to survive and even prosper when fitness landscapes shift. This summer we began the process of building a simulation environment, realistic robotic models, and the complex network of software necessary to study the evolution of populations of these robots across hundreds of generations. This summer’s work focused on the three core components of the software system, which will be optimized to run on Vassar’s multicore supercluster, Hopper. First, we designed and coded a genotype to phenotype (G-to-P) mapping algorithm that uses an unusual process-oriented (rather than part descriptive) mechanism to turn any randomly generated robot genotype into its corresponding phenotypic embodiment of sensors, motors, and neural network. This G-to-P mapping scheme is designed to allow simulation of gene duplication events as well as the mutation events typical in evolutionary robotics simulations. Next, we used the Gazebo simulation environment, which implements the Bullet physics engine, to construct a representation of the robot specified by the genome as well as the environment in which it must operate. Finally, we developed a fitness function that requires the solution to a behavior problem logically equivalent to the exclusive OR, a problem known to require neural networks of more than two layers to solve. Next steps will include the construction of the software infrastructure necessary to create a population of genotypes that will then be evaluated in the simulated environment and allowed to reproduce based on their success in that environment. A comparison of the evolvability of two populations, one evolved with and the other without the gene duplication mechanism turned on, will allow a clear test of our hypothesis.