Constrained Randomness and the Evolution of Artificial Neural Networks

• Activity Number: 186
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Thomas W. Woolley*+ and Steven F. Donaldson and Nick Dzugan and Jason Goebel
• Companies: Samford University and Samford University and Samford University and Samford University
• Keywords: constrained randomness ; lognormal ; evolution ; simulation

Abstract

Randomness is apparent in many aspects of evolution. None of its manifestations in nature, however, would be classified as pure chance or ontological randomness; rather randomness is always constrained. Data are generally characterized symmetrically with the arithmetic mean plus or minus either the standard deviation or the standard error of the mean. In reality many empirical distributions in the natural world are skewed making the assumption of symmetry improper and leading to possible misinterpretations of the data. Models that take into account the asymmetry of the data may be better served by underlying generative models such as the lognormal. Taking data asymmetry into account when describing it may lead to greater analytic quality and deeper insight into any information inherent in the data. The ultimate goal of our research is to see if it can be demonstrated that constraints on chance occurrences in evolution result in a set of boundary conditions that actually enable some level of predictability from a system generally viewed by scientists as incorporating any number of purely random features. This phase initiates our search for empirical evidence that supports (or refutes) the specification of quantitative factors underlying convergent evolution and/or self-organizing behavior in the natural world. With this in mind, we report on our phase I results, a thorough evaluation the distributional characteristics of one of our primary outcomes, time to reach target fitness (i.e., number of generations).