Robot Phylogenetics Kyle I. Harrington DEMO Lab Volen Center for Complex Systems Brandeis University Waltham, MA, USA kyleh@cs.brandeis.edu Jordan B. Pollack DEMO Lab Volen Center for Complex Systems Brandeis University Waltham, MA, USA pollack@cs.brandeis.edu ABSTRACT Bioinformatics techniques are introduced for the analysis of evolutionary search. These techniques are tested on build- able robots evolved in a virtual simulator for a locomotion task. By using bioinformatic visualizations properties of evolutionary search and relatedness between differing robot genotypes and phenotypes can be examined. Categories and Subject Descriptors I.2.9 [Computing Methodologies]: Artificial Intelligence— Robotics ; I.6.6 [Computing Methodologies]: Simulation and Modeling—Simulation Output Analysis General Terms Design, Experimentation, Measurement Keywords Evolutionary Algorithms, Robotics, Evolution, Behavior, Phy- logenetics, Evolutionary, Late Breaking Abstract The study of evolutionary algorithms (EAs) is often com- plicated by local optima, fitness plateaus, and other degen- erate behaviors. We introduce the use of bioinformatic vi- sualizations to gain insights into search spaces and trajecto- ries of EAs. Analysis techniques originally developed for the study of natural evolution are presented in the context of an artificial evolutionary substrate as robot phylogenetics. Evolutionary robotics involves searching a space of geno- types that map onto robot phenotypes. Successful examples include searching through spaces of robotic truss structures [4] and L-systems [2]. Our robots, evolved in breve, a 3D multi-agent simulator [3] and shown in Figures 1a & 1b, are similar to [6]. Some structural and all control parameters are evolved. A genetic algorithm with two-point crossover, and mutation (single-point randomization and +/- unit shift) is used on genomes of 18 genes. A population of size 100 is evaluated for 100 generations with tournament selection. Evolutionary computation has been used numerous times to solve bioinformatics problems [1]. However, bioinformat- ics techniques have not been used to solve evolutionary com- putation problems. We introduce the use multiple bioinfor- matics techniques to explore an EA. Heat maps of both the Copyright is held by the author/owner(s). GECCO’10, July 7–11, 2010, Portland, Oregon, USA. ACM 978-1-4503-0073-5/10/07. historically-ranked, Figures 1c and 1d, and fitness-ranked individuals, Figures 1e and 1f, show the evolutionary tra- jectory and the similarity in fitness landscapes, respectively. A cluster heat map [7] of genotype/phenotype observations, Figures 1g and 1h, suggests separate basins within the space and a close coupling between fitness and limb length. Pear- son’s correlation metric [5] clusters individuals which tend to be closer by minor mutation and intra-cluster crossover. Comparison of fitness landscapes suggest that the wave com- pression gene is the primary difference between optimal ra- dial and bilateral robot phenotypes. Finally, the long leg bias is made apparent by the fitness landscape. Bioinformatics techniques can elucidate analysis of evo- lutionary search. Clustering can provide insight into the roughness of a fitness landscape. Design of evolutionary sys- tems can be clarified by using bioinformatics visualization of evolutionary trajectories, spaces, and fitness landscapes. 1. ACKNOWLEDGEMENTS We thank Jon Klein for simulation advice, Lee Spector for his knowledge of the literature, Pengyu Hong for bioinfor- matics tips, and Bj¨ orn Gunnarsson for Walker discussions. 2. REFERENCES [1] G. Fogel and D. Corne. Evolutionary computation in bioinformatics. Morgan Kaufmann Pub, 2003. [2] G. Hornby and J. Pollack. Creating high-level components with a generative representation for body-brain evolution. Artificial Life, 8(3):223–246, 2002. [3] J. Klein and L. Spector. 3d multi-agent simulations in the breve simulation environment. Artificial Life Models in Software, pages 79–106. [4] H. Lipson and J. Pollack. Automatic design and manufacture of robotic lifeforms. Nature, 406(6799):974–978, 2000. [5] J. Rodgers and W. Nicewander. Thirteen ways to look at the correlation coefficient. American Statistician, pages 59–66, 1988. [6] L. Spector, J. Klein, K. Harrington, and R. Coppinger. Teaching the evolution of behavior with superduperwalker. In Proceedings of the 12th International Conference on Artificial Intelligence in Education. Citeseer, 2005. [7] L. Wilkinson and M. Friendly. The history of the cluster heat map. The American Statistician, 63(2):179–184, 2009.