Natural animals display awe-inspiring adaptability and robustness (e.g. three-legged dogs that can still run and catch frisbees). One reason neuroscientists believe animals are so capable is because their brains are structurally organized in that they exhibit neural regularity, modularity, and hierarchy. This project will test whether those neural properties also improve the robustness, adaptability, and overall intelligence of evolved computational agents, specifically robots. Evolutionary algorithms (EAs) evolve, rather than engineer, computational agents to produce qualities seen in nature, such as robustness and adaptability. While EAs often outperform human engineers, the robot bodies and brains they design pale in comparison to those of natural animals. Recent advances by the PI (Professor Clune) and others enable the evolution of neural networks (computational brains) that exhibit regularity, modularity, and hierarchy.
Professor Clune and his students will test whether regularity, modularity, and hierarchy, separately and in combination, improve (1) robustness to noise (2) robustness to damage (3) adaptability to new environments (4) learning, and (5) overall intelligence, measured as the ability to solve challenges of varying complexities. Based on his previous work and preliminary results, Clune anticipates that neural regularity, modularity, and hierarchy could increase all of these desirable behavioral qualities. Such knowledge will accelerate humanity's ability to deploy effective, autonomous robots, which will provide tremendous benefits to society (e.g. search and rescue, putting out fires, and elderly care). The research is woven into a tightly integrated educational plan that generates broader impacts via a robotics club in Laramie, a research-oriented course in evolutionary robotics, public education via videos and the press, and broadening participation in graduate training.