DYNAMIC SYSTEM PREDICTION USING TEMPORAL ARTIFICIAL NEURAL NETWORKS AND MULTI-OBJECTIVE GENETIC ALGORITHMS

摘要

We investigate the problem of learning to predict dynamical systems that exhibit switching behavior as a function of exogenous variables. The family of dynamical systems we present is significant to the modeling of gene expression and organismal response to environmental conditions and change. We first develop a framework for learning to predict events such as state or phase changes as a function of multiple dynamic variables. Next, we consider the more challenging problem of identifying parameters and the functional form of the dynamical systems ab initio. We then survey several applicable representations and inductive learning techniques for each task. We then describe a comparative experiment in learning a particular instantiation of the dynamical system for a plant genome modeling application. Finally, we evaluate the results using predictive accuracy of the differential equation parameters or accuracy on the event prediction task; consider the ramifications for modeling the metabolic processes of living systems; and outline future challenges such as multi-objective optimization and finding relevant exogenous and latent variables.