A1191
Title: Genetic algorithms in regression
Authors: Mo Li - UL at Lafayette (United States) [presenting]
Qiqi Lu - Virginia Commonwealth University (United States)
Robert Lund - The University of California, Santa Cruz (United States)
Xueheng Shi - University of Nebraska-Lincoln (United States)
Abstract: Many statistical problems require optimization over a discrete parameter space with unknown and potentially varying dimensions. In these settings, gradient based methods are often ineffective due to non-differentiability and the combinatorial growth of candidate models. GAReg is introduced, a unified genetic-algorithm package for discrete optimization in regression problems where standard algorithms are impractical or unjustified. GAReg uses a compact chromosome representation that supports a range of tasks, including optimal knot placement for regression splines, best-subset variable selection, multiple changepoint detection, and related discrete model search problems. The package implements uniform initialization, constraint preserving crossover and mutation operators, steady state replacement, and optional island model parallelization. By enabling efficient exploration of high-dimensional model spaces, GAReg delivers near optimal solutions in regimes where exhaustive enumeration, integer programming, or dynamic programming approaches are computationally infeasible.
