Skip to main content
2011-2012 Colloquium Series: Michael Palmer


Evolved Neurogenesis and Synaptogenesis for Robotic Control
11am Friday, January 13th, 2012

International Center (CIP-115)

We have developed a novel system to "grow" neural networks according to an inherited set of production rules (the genotype), inspired by indenmayer systems. In the first phase, the neurons proliferate in three-dimensional space by cell division, and differentiate in function, according to the production rules. In the second phase, axons emerge from the neurons and seek out connection targets. Part of each production rule is an arithmetic expression that is evaluated to facilitate decisions during both phases of growth. A feed-forward neural network is thus produced. We connect each network to a (fixed) robotic body with a set of input sensors and muscle actuators. The robot is placed in a physically simulated environment and controlled by its network for a certain time, receiving a fitness score according to its behavior (the phenotype). Mutations are introduced into offspring by making changes to their sets of production rules; this may produce new networks, with new (and sometimes adaptive) behaviors. Our first evolutionary experiments with the system have produced controllers for 18-degree-of-freedom robotic "spiders" with the ability to "gallop", and to follow a compass heading, according to their sensor inputs. (Videos ere: ) The long-term goal of this work is the understanding of "scalable" methods to evolve neural etworks for increasingly complex control tasks.

Bio: Dr. Michael Palmer is a Research Associate in the laboratory of Marc Feldman in the Biology Department at Stanford University. He received his B.S. in Physics from Yale, and his Ph.D. in Computer Science from Caltech. In his research, he studies evolvability and macroevolution, using in silico models. One half of his work is theoretical, using biologically-inspired models like lattice proteins and gene networks, to ask how evolution accretes complexity over long time periods. The other half of his work is applied, asking what is required, in practice, to "scale up" the complexity of evolved robotic controllers.