Jeff Clune (Director, Evolving Artificial Intelligence Laboratory)
Abstract. This talk summarizes two different, but related research projects. (1) Evolving Neural Networks that are Modular, Hierarchical, and Regular: A major open question in biology is how evolution produced the complexity seen in nature. I study this question by attempting to evolve such complexity in computational simulations of evolution, specifically in computational brain models called neural networks. To date, neural network phenotypes evolved in silico tend to be simple and lack much of the structural organization found in animal brains. I will describe new approaches that combine computational abstractions of evolution and developmental biology to automatically produce neural networks that are modular, hierarchical, and regular. Moreover, I will show that these properties increase the speed of evolution and level of intelligence it generates. (2) Robots that Recover from Damage: While animals can quickly adapt to injuries, current robots cannot "think outside the box" to find a compensatory behavior when they are damaged. I will introduce an intelligent trial and error algorithm that allows robots to adapt to damage in less than two minutes. Before deployment, a robot exploits a novel algorithm to create a detailed map of the space of high-performing behaviors: This map represents the robot's intuitions about what behaviors it can perform and their value. If the robot is damaged, it uses these intuitions to guide a trial-and-error learning algorithm that conducts intelligent experiments to rapidly discover a compensatory behavior that works in spite of the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways.