Introduction
Dynamic Multi-objective Optimization Problems (DMOPs) are prevalent in real-world scenarios where objectives and constraints change over time, such as in dynamic logistics or fluctuating resource allocation. The primary challenge in DMOPs is maintaining efficiency; when the environment shifts, algorithms must quickly re-converge to the new Pareto-Optimal Front (POF). Traditional evolutionary algorithms often struggle to adapt rapidly without significant computational overhead or manual tuning.
Methodology: The CORAL-IPG Framework
To address these challenges, the authors propose CORAL-DMOEA, a framework that leverages Correlation Alignment (CORAL) to transfer knowledge from previous environments to the current one. The core of this framework is the CORAL-IPG (Initial Population Generator) sub-algorithm.
The process involves three key steps:
- Covariance Matching: The algorithm calculates the covariance of the distribution of successful solutions in the previous environment.
- Alignment: A linear transformation is applied to these solutions to align their distribution with the estimated distribution of the new environment.
- Warm-Starting: These aligned solutions form the initial population for the next evolutionary cycle, providing a high-quality starting point.
Unlike parametric transfer learning methods, CORAL is non-parametric and computationally efficient, requiring no manual hyperparameter tuning, making it a robust "plug-and-play" solution.
Experimental Evaluation
The framework was tested using the IEEE CEC 2015 benchmark suite, which includes complex functions like FDA4, FDA5, and the HE series. Performance was evaluated using the Modified Inverted Generational Distance (MIGD) and Modified Hypervolume (MHV) metrics across various change frequencies (ranging from 25 to 200 generations).
Key Results and Discussion
The results demonstrate that CORAL-DMOEA significantly enhances the performance of standard algorithms like MOEA/D and NSGA-II. The Ratio of Change (ROC) metric was used to quantify improvements, where positive values indicate superior performance over the baseline.
Key findings include:
- High Effectiveness: CORAL-enhanced versions outperformed baselines in 87.5% of cases for MIGD and 91.67% for MHV.
- Rapid Convergence: In benchmarks like FDA4 and HE2, the algorithm achieved dramatic speedups, finding the new optimal front almost immediately after environmental changes.
- Robustness: The method proved versatile across different types of environmental dynamics. However, it faced challenges with "isolated" functions (e.g., FDA5_iso) where the statistical relationship between successive environments is too weak for simple correlation alignment to bridge.
Conclusion
CORAL-DMOEA offers a robust solution for dynamic optimization by effectively recycling past knowledge. By aligning the statistical properties of data across time steps, it provides a significant head start for evolutionary solvers in changing environments. Future work will focus on applying this framework to complex, real-world engineering challenges where dynamic optimization is a necessity.
Summary updated