← All Research Themes
Learning-driven Evolutionary Optimization
Overview
Integrating learning and transfer into evolutionary search to improve adaptation, convergence, and robustness in dynamic and high-dimensional optimization.
Key Questions
- How can learning mechanisms be integrated into evolutionary algorithms without introducing unnecessary computational overhead?
- What kinds of knowledge should be transferred between optimization phases, and how can negative transfer be avoided?
- How can learned models generalize across problem instances and changing environments?
Our Contributions
- Developed transfer-learning frameworks for dynamic multiobjective optimization that reuse historical knowledge to accelerate convergence
- Introduced knee point-based imbalanced transfer learning to handle distribution mismatch in changing environments
- Proposed spatial-temporal knowledge transfer for dynamic constrained optimization problems
Recommended Papers
Introduces transfer learning for dynamic multiobjective optimization, so search can reuse prior environments.
Uses knee-point-aware transfer to avoid negative transfer when fronts shift unevenly.
Recent Work
Recasts dynamic multiobjective optimization as a prompt-guided time-series prediction problem and shows that large language models can generate competitive solutions.
Open Problems
- Developing theoretically grounded transfer criteria that can determine when and what to transfer
- Scaling learning-driven approaches to high-dimensional and many-objective problems
- Understanding the interaction between learning mechanisms and population diversity
Prospective Student Projects
- Transfer learning for many-objective optimization in dynamic environments
- Meta-learning approaches for rapid adaptation to new optimization problems
- Learning-based constraint handling in dynamic constrained optimization