A Data Stream Ensemble Assisted Multifactorial Evolutionary Algorithm for Offline Data-Driven Dynamic Optimization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Optimization

Background:

• Existing offline data-driven optimization primarily addresses static environments.
• Dynamic environments pose challenges due to time-varying data distributions, necessitating adaptive surrogate models and solution tracking.

Purpose of the Study:

• To propose a knowledge-transfer-based data-driven optimization algorithm for dynamic environments.
• To enable effective tracking of optimal solutions in environments with evolving data distributions.

Main Methods:

• An ensemble learning approach trains surrogate models, incorporating historical and new environmental data.
• Historical models are retrained with new data, forming base learners for an ensemble surrogate model.
• Base learners and the ensemble model are optimized simultaneously in a multi-task environment.

Main Results:

• The proposed algorithm demonstrates effectiveness on six dynamic optimization benchmark problems.
• Empirical results show superior performance compared to four state-of-the-art offline data-driven optimization algorithms.
• The knowledge-transfer mechanism accelerates the tracking of optima in current environments.

Conclusions:

• The knowledge-transfer-based ensemble learning algorithm effectively addresses offline data-driven optimization in dynamic environments.
• The method successfully leverages past knowledge to adapt to and optimize in changing conditions.
• The algorithm provides a robust solution for tracking optimal solutions in time-varying optimization problems.