Nonlinear model structure design and construction using orthogonal least squares and D-optimality design.

Area of Science:

• Statistics
• Machine Learning
• Computational Science

Background:

• Model subset selection is crucial for developing accurate and reliable predictive models.
• Existing methods may not adequately balance model approximation ability with robustness and adequacy.
• Efficient algorithms are needed for complex datasets and high-dimensional modeling.

Purpose of the Study:

• To introduce a novel, efficient learning algorithm for model subset selection.
• To develop a composite cost function optimizing model approximation, robustness, and adequacy.
• To ensure parsimony and reliability in the selected model subsets.

Main Methods:

• Utilizing forward orthogonal least squares (OLS) for parameter estimation.
• Incorporating a D-optimality design criterion within the cost function.
• Constructing a D-optimality-based cost function leveraging the orthogonalization process for computational efficiency.

Main Results:

• The proposed algorithm efficiently selects optimal model subsets.
• The composite cost function successfully balances approximation ability with robustness and adequacy.
• The D-optimality criterion enhances model robustness, adequacy, and parsimony.

Conclusions:

• The new algorithm provides an effective and computationally efficient approach to model subset selection.
• The integration of D-optimality significantly improves the quality and reliability of selected models.
• This method offers a valuable tool for statistical modeling and machine learning applications.