Uncertainty-Aware Offline Data-Driven Multi-Objective Optimization
Researchers solve dominance errors in data-limited optimization with flexible dual-ranking.
In offline data-driven multi-objective optimization (MOO), surrogate models trained solely on static datasets introduce epistemic uncertainty, leading to incorrect dominance judgments that misguide search processes. Existing methods, which rely on Gaussian Process Regression (GPR) for uncertainty estimates, are computationally expensive and cannot leverage other uncertainty quantification techniques.
To address this, Huanbo Lyu and nine co-authors from multiple institutions introduce a simple yet effective dual-ranking strategy that flexibly integrates predictive results and uncertainty estimates from any surrogate model. The approach performs non-dominated sorting using both surrogate-based fitness values and uncertainty-aware fitness values, prioritizing candidate solutions that are both high-quality and reliable. Extensive experiments, including ablation, sensitivity, and comparative studies, demonstrate the strategy's effectiveness and robustness across different surrogates, making it highly suitable for data-limited real-world applications.
- Dual-ranking strategy uses both predictive and uncertainty-aware fitness values for non-dominated sorting.
- Works with any surrogate model, not just GPR, overcoming computational cost and scalability limitations.
- Experimental validation shows enhanced robustness and solution quality in data-constrained environments.
Why It Matters
Enables more reliable decision-making in engineering and science where data is scarce and uncertainty high.