QUIVER: Cost-Aware Adaptive Preference Querying in Surrogate-Assisted Evolutionary Multi-Objective Optimization
Adaptive optimization cuts regret by 25% on complex WFG problems.
Interactive multi-objective optimization faces a budget dilemma: should resources go toward expensive objective evaluations or toward eliciting decision-maker preferences? Existing methods pick one modality—either cheap but noisy pairwise statements (PS) or costlier indifference adjustments (IA). QUIVER, presented at GECCO '26, introduces a cost-aware framework that dynamically selects the best next action—objective evaluation, PS, or IA—by maximizing expected decision-quality improvement per unit cost.
Tested on DTLZ and WFG benchmarks with synthetic decision-makers, QUIVER consistently outperforms single-modality baselines. On hard WFG4 problems, it achieves utility regret of 2.14; on WFG9, 2.82—a 25% improvement. The adaptive modality selection is striking: on easy DTLZ2, it uses 80% PS queries; on hard WFG9, it shifts to 35% IA queries, showing cost-aware preference learning in action. This work opens the door to more efficient, human-in-the-loop optimization systems.
- QUIVER adaptively chooses between objective evaluations and two preference query types (pairwise statements vs. indifference adjustments) based on cost-per-unit-quality improvement.
- Achieves utility regret of 2.14 on WFG4 and 2.82 on WFG9, a 25% improvement over single-modality baselines.
- Demonstrates dynamic modality selection: 80% cheap PS queries on easy DTLZ2, shifting to 35% richer IA queries on hard WFG9 problems.
Why It Matters
Makes multi-objective optimization cheaper and smarter by spending resources where they yield the most decision insight.