On the effect of response transformations in sequential parameter optimization.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning

Background:

• Parameter tuning is crucial for evolutionary algorithms (EAs).
• The sequential parameter optimization (SPO) framework is a model-assisted approach for tuning stochastic optimizers.
• Established algorithms exist within the SPO framework.

Purpose of the Study:

• To enhance the SPO framework by incorporating transformation steps.
• To empirically analyze the impact of different transformations on EA parameter tuning.
• To investigate improvements in model properties and discriminatory power.

Main Methods:

• Integration of transformation steps (e.g., rank, Box-Cox) before response aggregation and modeling.
• Application of design-of-experiments techniques for empirical analysis.
• Analysis of resulting models, residual distributions, and model-based effect plots.

Main Results:

• Rank transformation of responses significantly improves SPO framework performance.
• Rank and Box-Cox transformations enhance residual distribution symmetry and normality.
• Model-based effect plots demonstrate increased discriminatory power with rank transformation.

Conclusions:

• Transformations, especially rank transformation, offer substantial benefits to the SPO framework for EA parameter tuning.
• The enhanced SPO framework leads to more robust and effective optimization algorithms.
• Further investigation into adaptive procedures and transformations is warranted.