RI: Small: RUI: Synthesis of Robust Artificial Systems by Adaptive Genetic Programming

Hampshire College, Amherst MA

Investigators

Abstract

Computer-based problem-solving systems are revolutionizing many areas of science and engineering, with pervasive impacts on economic activity, human health, national security, and the advancement of science. Several of the most powerful and promising approaches to the development of these systems borrow ideas from biology, for example, when artificial neural networks are used to enable computer systems to learn. The processes of random variation and fitness-based selection motivated by biology have been particularly useful in several applications, but they have not yet produced the kind of radical innovations that are characteristic of living systems. In this project, key elements of genetic programming, such as the processes governing variation, will themselves be allowed to adapt, with the aim of producing more powerful problem-solving computer systems. These systems may have applications in several areas of science and engineering. The project will be conducted in the context of educational activities that integrate research and education across undergraduate and graduate levels, thereby providing training to a new generation of computational scientists. The primary goal of the proposed project is to enhance genetic programming technologies in ways that will allow them to more routinely produce more innovative solutions to difficult problems, and to produce systems that perform well in complex environments. The central hypothesis underlying this effort is that the innovating power of biology, and the power of biology to produce robust systems, stems in part from the fact that the adaptive mechanisms of biology themselves adapt. Self-adaptive genetic programming systems, in which the algorithms for variation and selection are themselves subject to variation and selection, have been explored for decades but have only recently begun to show practical promise for solving difficult problems. The proposed project will begin with a promising system of this type and will test it systematically, in order to elucidate general principles that will then be used to develop and apply more refined, adaptive algorithms. Applications ranging from the automatic programming of exercises in a first-semester programming textbook to the development of multicellular organisms in a virtual ecosystem will be used to test and demonstrate the systems developed in this project.