Minimization of Akaike's Information Criterion in Linear Regression Analysis via Mixed Integer Nonlinear Program

TL;DR

This paper introduces a novel branch-and-bound algorithm for directly minimizing Akaike's Information Criterion (AIC) in linear regression models, improving model selection accuracy especially for small to medium datasets.

Contribution

It develops a mixed integer nonlinear programming approach with specific bounds and branching strategies, integrated into SCIP, to find optimal models based on AIC.

Findings

Successfully identifies best models for small and medium datasets.

Finds high-quality solutions for large datasets.

Outperforms traditional stepwise methods in accuracy.

Abstract

Akaike's information criterion (AIC) is a measure of the quality of a statistical model for a given set of data. We can determine the best statistical model for a particular data set by the minimization of the AIC. Since we need to evaluate exponentially many candidates of the model by the minimization of the AIC, the minimization is unreasonable. Instead, stepwise methods, which are local search algorithms, are commonly used to find a better statistical model though it may not be the best. We propose a branch and bound search algorithm for a mixed integer nonlinear programming formulation of the AIC minimization by Miyashiro and Takano (2015). More concretely, we propose methods to find lower and upper bounds, and branching rules for this minimization. We then combine them with SCIP, which is a mathematical optimization software and a branch-and-bound framework. We show that the proposed method can provide the best statistical model based on AIC for small-sized or medium-sized benchmark data sets in UCI Machine Learning Repository. Furthermore, we show that this method finds good quality solutions for large-sized benchmark data sets.