System Neural Network: Evolution and Change Based Structure Learning

Abstract

System evolution analytics with artificial neural networks is a challenging and path-breaking direction, which could ease intelligent processes for systems that evolve over time. In this article, we contribute an approach to do <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Evolution and Change Learning</i> (ECL), which uses an evolution representor and forms a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">System Neural Network</i> (SysNN). We proposed an algorithm <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">System Structure Learning</i> , which is divided in two steps. First step uses the evolution representor as an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Evolving Design Structure Matrix</i> (EDSM) for intelligent design learning. Second step uses a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Deep Evolution Learner</i> that learns from evolution and changes patterns of an EDSM to generate Deep SysNN. The result demonstrates application of the proposed approach to analyze four real-world system domains: software, natural-language, retail market, and movie genre. We achieved significant learning over highly imbalanced datasets. The learning from previous states formed SysNN as a feed-forward neural network, and then memorized information as an output matrix to recommend entity-connections.