Optimal additions to asymmetrical fractional factorial designs under Lee discrepancy

Chatterjee Kashinath, Gou Tingxun, Hong QinScientia 2020 · DOI

Abstract

Due to the character that the sequential experiment enlarges stage by stage, the new runs will be added into the previous design that has been chosen. How to design a good sequential experiment becomes an interesting issue. This paper studies the uniform asymmetric extended designs in terms of Lee discrepancy for a possible application to sequential experimentation. An analytical relation between Lee discrepancy of two- and three-level mixed extended designs, and its initial and additional designs is established, which provides a statistical justification to construct (nearly) uniform extended designs. As a benchmark for choosing best extended designs, a lower bound of Lee discrepancy of extended designs is obtained. In addition, some examples are also presented to further support our theoretical findings.

Introduction

In industrial and pharmaceutical research and development, experiments are frequently conducted sequentially. Rather than a single large-scale trial, researchers often begin with an initial design and subsequently add more runs to refine their results or increase precision. The fundamental challenge in this approach is ensuring that the combined set of trials—the initial design plus the additions—remains "uniform."

A uniform design distributes experimental points as evenly as possible across the design space. This property is vital for capturing complex, non-linear system behaviors without relying on rigid model assumptions. While previous research has largely focused on symmetric designs where all factors have the same number of levels, this paper addresses the more complex case of asymmetrical (mixed-level) designs, specifically those involving factors with two and three levels.

Methodology: The Lee Discrepancy Framework

The authors employ Lee discrepancy as the primary metric for evaluating design uniformity. Unlike model-dependent metrics, Lee discrepancy measures how effectively experimental points fill a multi-dimensional space.

Analytical Decomposition

A core contribution of this work is the derivation of an analytical relationship (Theorem 3.1) that decomposes the total Lee discrepancy of an extended design. This total is expressed as a function of:

  1. The discrepancy of the initial design.
  2. The discrepancy of the additional runs.
  3. A "cross-term" representing the interaction between the initial and additional designs.

This decomposition allows researchers to mathematically justify the selection of specific additions rather than relying on trial and error.

Theoretical Lower Bounds

To provide a benchmark for optimality, the authors derived theoretical lower bounds for the Lee discrepancy of these mixed-level designs. These bounds act as a "gold standard," enabling researchers to quantify how close a specific design is to achieving perfect uniformity.

Search Algorithm

The paper introduces a stochastic search algorithm (Algorithm 1) designed to find nearly optimal sets of additional trials. By utilizing the mathematical properties of Kronecker products and "meeting numbers" (which count level matches between runs), the algorithm efficiently navigates the vast search space of potential design matrices.

Key Insights and Results

The research demonstrates that the quality of an extended design is not merely the sum of its parts. The "cross-term" is critical; a high-quality initial design combined with a high-quality additional design does not automatically result in a high-quality combined design unless their interaction is optimized.

Numerical examples provided in the paper, using 12-run and 18-run initial designs, show that the proposed search algorithm can identify additions that reach 100% efficiency (matching the theoretical lower bound) for small additions. For larger, more complex additions, the algorithm consistently maintains efficiencies above 94%.

Furthermore, this method offers significantly more flexibility than traditional "fold-over" designs. While fold-over methods typically require doubling the number of experimental runs, the Lee discrepancy approach allows for adding a small, specific number of runs while maintaining high uniformity.

Conclusion and Future Directions

This work provides a rigorous mathematical foundation for growing asymmetrical experimental designs. By establishing analytical formulas and lower bounds for the 2-3 mixed-level case, the authors have created a practical tool for researchers in high-tech engineering and manufacturing.

Future work may involve extending these results to designs with four or more levels, which would further increase the complexity of the search space. Additionally, exploring multi-objective optimization that combines Lee discrepancy with other criteria, such as orthogonality, could lead to even more robust experimental frameworks.

Summary updated

Abstract

Due to the character that the sequential experiment enlarges stage by stage, the new runs will be added into the previous design that has been chosen. How to design a good sequential experiment becomes an interesting issue. This paper studies the uniform asymmetric extended designs in terms of Lee discrepancy for a possible application to sequential experimentation. An analytical relation between Lee discrepancy of two- and three-level mixed extended designs, and its initial andadditional designs is established, which provides a statistical justification to construct (nearly)uniform extended designs. As a benchmark for choosing best extended designs, a lowerbound of Lee discrepancy of extended designs is obtained. In addition, some examples are alsopresented to further support our theoretical findings.
BibTeX
@inproceedings{kashinath2020optimal,
  author = {Chatterjee Kashinath and Gou Tingxun and Hong Qin},
  title = {{Optimal additions to asymmetrical fractional factorial designs under Lee discrepancy}},
  booktitle = {Scientia},
  year = {2020},
  doi = {10.1360/n012019-00095},
}
Figure