MOTA: A Many-Objective Tuning Algorithm Specialized for Tuning under Multiple Objective Function Evaluation Budgets.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Algorithm Tuning

Background:

• Selecting appropriate parameter values is crucial for effective optimization algorithm performance.
• Tuning stochastic algorithms presents challenges due to noise and multiple performance objectives.
• Existing methods may not adequately address the complexities of many-objective tuning.

Purpose of the Study:

• To introduce a novel many-objective tuning algorithm (MOTA) for stochastic optimization algorithms.
• To develop a specialized approach for optimizing multiple performance measures across diverse evaluation budgets.
• To enhance the effectiveness and efficiency of parameter selection for optimization algorithms.

Main Methods:

• Formulation of a tuning problem with speed objectives and decision variables.
• A control parameter tuple assessment using single run history across multiple budgets.
• Implementation of a preemptively terminating resampling strategy to handle noise.
• Application of bi-objective decomposition for many-objective optimization.
• Integration with differential evolution operators for parameter search.

Main Results:

• The Many-Objective Tuning Algorithm (MOTA) was developed and tested.
• MOTA demonstrated effectiveness in tuning algorithms like NSGA-II and MOEA/D.
• The algorithm successfully optimized multiple performance measures over various evaluation budgets.
• The specialized components of MOTA contributed to its successful application.

Conclusions:

• MOTA provides an effective solution for the many-objective tuning of stochastic optimization algorithms.
• The proposed methods address key challenges in tuning, including noise and multi-objective performance.
• MOTA enhances the ability of practitioners to select suitable parameter values for their specific problems.