This action funds an NSF Postdoctoral Research Fellowship for FY 2010. The fellowship supports a research and training plan entitled "Identifying the Environmental Attributes that Drive the Evolution of Complexity in Organisms" for Jeff Clune. The host institution for this research is Cornell University, and the sponsoring scientist is Lipson Hod.
One of the major open questions in biology is how evolution produces the complexity seen in natural forms, such as the structural organization (regularity, modularity, and hierarchy) of bodies and brains. The environmental pressures that drive the evolution of such structural organization are poorly understood largely because they are difficult to study experimentally in natural systems. Understanding structural organization is important, however, because it is a fundamental property of biological systems, because it affects how organisms will respond to novel environments, and because it increases the speed of evolutionary adaptation, which opens possibilities for improving bioengineering. The Fellow is taking advantage of computational systems that exhibit evolutionary dynamics to identify the environmental attributes that drive the evolution of structural organization. Specifically, the research (1) develops novel statistical algorithms to quantify structural organization, (2) uses evolution experiments in computers to test whether eight different environmental attributes drive the evolution of structural organization, (3) quantifies structural organization in biological phenotypes, (4) tests if the environmental drivers of structural organization in computer evolution experiments (identified in 2) also drive structural organization in biological phenotypes (using data from 3).
The training goals include enhancing mathematical and machine learning skills and better understanding computational models of brains. The broader impacts include development of a website that allows visitor participation in evolving structurally-organized phenotypes, and publishing open-source software tools for quantifying structural organization.
Identifying the Environmental Attributes that Drive the Evolution of Structural Organization. By Jeff Clune (JeffClune.com) Intellectual Merit: One of the major open questions in biology is how evolution produced the complexity seen in natural forms. The bodies and brains of natural organisms are complex in part because they exhibit structural organization, meaning that they are modular, regular, and hierarchical. The environmental pressures that drive the evolution of such structural organization are poorly understood, however, largely because they are difficult to experimentally study in natural systems. Therefore, researchers have turned to computational systems that exhibit evolutionary dynamics, but to date the evolved phenotypes of digital organisms tend to be simple, non-modular, and irregular. Our inability to create synthetic evolutionary processes that produce phenotypes with structures nearly as organized as natural phenotypes reveals how little we know about the evolutionary origins of structural organization. Understanding structural organization is important, however, because it is a fundamental property of biological systems, because it affects how organisms will respond to novel environments, and because it increases evolvability, which can improve bioengineering practices. Research Discoveries & Outcomes: My fellowship resulted in many findings. Initially, my colleagues and I discovered a simple, yet powerful force that causes the evolution of ?modular networks. The evolution of modularity is a central biological question, and is ?related to one of the longest-standing open biological questions of why populations are ?so evolvable (i.e. able to quickly adapt to new environments). Our result will thus have immediate implications for a wide-variety of fields, including neuroscience, genetics and harnessing evolution for engineering purposes. A paper summarizing this work is available (free to the public) on the ArXiv preprint server (http://arxiv.org/abs/1207.2743v1) and is under review at a general-audience science journal. A second discovery was the different types of network properties that evolve with ?different generative encodings, which are ways of storing information in a genotype and translating that information into a phenotype. My colleagues and I found that encodings based on iterative rewriting combined with embryonic geometrical patterning produce networks that evolve to be scale-free and modular. This discovery was published in the peer-reviewed Proceedings ?of the Genetic and Evolutionary Computation Conference, 2011. A third discovery was that evolving three-dimensional morphologies with a particular abstraction of biological development can produce morphologies that share many properties with natural forms, ?such as symmetries and repetitions, with and without variation. This work suggests that ?one of the key complexity-generating forces in natural development is the composition ?of geometric coordinate frames similar to the geometric patterns seen in developing ?embryos. This work was published in the peer-reviewed Proceedings of the European Conference on Artificial Life, 2011. A website EndlessForms.com was created so that the public could participate in this science project by breeding objects generated with this encoding (see Broader Impacts below). A final result is that these concepts can solve hard engineering problems. My colleagues and I evolved gaits for physical robots with these techniques and the resultant gaits were better than those produced by humans and traditional machine learning algorithms. This work was ?published in the peer-reviewed Proceedings of the European Conference on Artificial Life, 2011. A follow-up study combining these techniques with a simulator will be submitted for publication this month (October, 2012). We further created a new robot design that is better for testing robotic gaits, which was published in the peer-reviewed 2012 Artificial Life Conference. The robotic design and software are open-source and freely available. Training: During my fellowship, I learned (1) how to quantitatively measure modularity in networks, (2) deep learning techniques that allow neural networks to learn to recognize the objects in their world, (3) how to use new web technologies for developing 3D websites, specifically Django, WebGL, Nginx, Javascript, CSS, Amazon’s EC2, and (4) 3D printing techniques. Broader impacts: In addition to my publications, I helped create EndlessForms.com, a website that allows the general public to learn about evolution while evolving three-dimensional phenotypes that can be 3D printed. To date, nearly 3.5 million objects have been evaluated on the site from approximately 50,000 unique visitors from over 150 countries. There has been press coverage of the site by MSNBC, the New Scientist, MIT's Technology Review, Slashdot, and over 40 other outlets. EndlessForms was rated as one of the top 35 websites in the 3D printing industry, was chosen as a finalist for the Evolutionary Art Competition at the 2011 Genetic and Evolutionary Computation Conference, and a video about EndlessForms was a finalist in the 2012 AAAI Video competition. A journal paper summarizing EndlessForms and utilizing the database of 3.5 million rated objects is currently being prepared. I also assisted with the training of five graduate and five undergraduate students, including publishing papers with three female engineers, who are significantly underrepresented in engineering.
