Solving Dynamic Multiobjective Problem via Autoencoding Evolutionary Search.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Evolutionary computation

Background:

• Dynamic multiobjective optimization problems (DMOPs) are prevalent in real-world scenarios and have garnered significant research interest.
• Existing methods for solving DMOPs often face challenges in efficiently tracking changes in Pareto-optimal solutions.

Purpose of the Study:

• To propose a novel autoencoding evolutionary search method for effectively solving DMOPs.
• To enhance the tracking of dynamic changes in Pareto-optimal solutions using an autoencoder-based prediction approach.

Main Methods:

• An autoencoder is developed to predict the movement of Pareto-optimal solutions by learning from previously obtained nondominated solutions.
• This autoencoder is integrated into established multiobjective evolutionary algorithms (EAs) like NSGA-II and MOEA/D.
• The prediction method offers a closed-form solution, minimizing computational overhead during evolutionary search.

Main Results:

• Comprehensive empirical studies demonstrate the efficacy of the proposed autoencoding evolutionary search.
• The method shows competitive performance compared to three state-of-the-art prediction-based dynamic multiobjective EAs on standard benchmarks.
• The autoencoder's ability to learn from ongoing optimization processes leads to more accurate Pareto-optimal solution predictions.

Conclusions:

• The proposed autoencoding evolutionary search is an effective approach for solving DMOPs.
• The integration of autoencoders provides an efficient and accurate prediction mechanism for dynamic changes in optimization problems.
• This method offers a promising direction for advancing research in dynamic multiobjective optimization.