Session:

GSW - Graduate Student Workshop

Title:

Time Series Prediction by Genetic Programming with Relaxed Assumptions in Mathematica

Authors:

Stuart W. Card

Abstract:

Time series produced by black box systems with both stochastic and nonlinear dynamical components have proven resistant to prediction. Also, prediction alone is unsatisfying: insight into the hidden dynamics is desired. Automatic induction of a system model would be ideal. A genetic programming (GP) / neural network (NN) / wavelet approach is motivated. An initial test problem selection is justified. Data preprocessing is described. The GP is shown to rely on weaker assumptions than those implicit in orthodox methods. An implementation in Mathematica is illustrated. GP discovery of equations, NN optimization of their parameters, and joint time-frequency representations, should provide highly parsimonious descriptions, capturing local and global characteristics of stochastic attractors, amenable to meaningful interpretation.