Evolutionary Robotics

We are applying computational models of evolution to the creation of software control systems for coordinated teams of robots. The goal is to create intelligent, dynamically autonomous systems. Our current focus is on a variety of open issues, including characterizing the power and tradeoffs associated with alternative behavioral representations, bridging the gap between simulation and the real world with continuous and embedded learning, and the application of coevolution to morphology as well as behaviors.

Coevolutionary Models

To successfully apply evolutionary algorithms to the creation of behaviors for teams of intelligent, dynamically autonomous robots, we must develop effective techniques for evolving solutions in the form of interacting coadapted components. We have developed a general architecture for evolving such components as a collection of two or more genetically-isolated species that interact and adapt to one another within a shared domain model. He are interested in harnessing the game-theoretic power of coevolution to learn team behaviors, while at the same time understanding the effects certain key design choices and problem properties have on successful performance.

Representation Issues

Evolving software control systems for coordinated teams of autonomous robots will require suitable behavioral representations. These can be difficult to design since they must be both specialized and easily evolvable. Existing theoretical work in evolutionary biology may help provide us with a principled way of designing specialized representations and reproductive operators. By building on this work, we intend to develop principles for designing representations and operators for the next generation of evolutionary algorithms.

Continuous and Embedded Learning (Anytime Learning)

Intelligent, dynamic autonomy will require systems that can adapt to changes in the environment and to changes in their own capabilities, as well as to changes within the team in which they are a member. We are investigating several hybrid machine learning techniques, and various representations, as well as designing a system that allows for safely learning and adapting while in the real world.