Learning Multimodal Parameters: A Bare-Bones Niching Differential Evolution Approach.

Area of Science:

• Computational Intelligence
• Machine Learning
• Optimization

Background:

• Optimization is a critical substep in most machine learning methods.
• Nondifferentiability and multimodality of objectives necessitate advanced optimization techniques.
• Evolutionary multimodal optimization (MMOP) enables simultaneous learning of diverse model parameters.

Purpose of the Study:

• To develop an efficient neighborhood-based niching algorithm for locating multiple optima in multimodal landscapes.
• To enhance the performance of evolutionary multimodal optimization (MMOP) for applications requiring accuracy and diversity.

Main Methods:

• A novel neighborhood-based niching algorithm is developed, using bare-bones differential evolution as a baseline.
• Gaussian mutation with local mean and standard deviations is employed to capture landscape niches.
• A diversity-preserving operator is integrated to reinitialize converged or overlapped neighborhoods, enhancing global exploration.

Main Results:

• The proposed algorithm demonstrates superior and consistent performance across a wide range of MMOP problems.
• Experimental results validate the algorithm's effectiveness in locating multiple optima simultaneously.
• Successful application in training neural network ensembles confirms its practical benefits for learning multimodal parameters.

Conclusions:

• The developed neighborhood-based niching algorithm is an efficient solution for evolutionary multimodal optimization.
• The method offers significant advantages for machine learning applications demanding both accuracy and parameter diversity.
• This work advances the field of MMOP by providing a robust tool for complex optimization landscapes.