Knowledge Gradient for Multi-Objective Bayesian Optimization with Decoupled Evaluations

2023-02-02

AI Summary

In practical multi-objective Bayesian optimization (MOBO), objectives often exhibit asynchronous and decoupled evaluation—i.e., differing evaluation latencies, costs, and availability—yet conventional MOBO methods assume synchronous, coupled evaluations. Method: This paper introduces the first decoupled-evaluation-aware multi-objective knowledge gradient (MO-KG) framework. It explicitly models inter-objective evaluation delays and cost heterogeneity, designs a target-adaptive sampling policy, and enables efficient MO-KG computation via Gaussian process surrogates and Monte Carlo gradient estimation. Contribution/Results: Theoretically, the Pareto front estimator is proven asymptotically consistent. Empirically, on standard benchmarks, the method reduces average total evaluations by up to 37% while significantly improving Pareto coverage and hypervolume (HV). This work constitutes the first generalization of the knowledge gradient to decoupled multi-objective settings, jointly optimizing evaluation cost and Pareto front convergence.

Problem

• Optimizes multi-objective trade-offs with minimal samples.
• Handles decoupled objectives with varying evaluation costs.
• Proposes a cost-aware acquisition function for faster learning.

Innovation

• Scalarization-based knowledge gradient acquisition function
• Decoupled evaluations for multi-objective optimization
• Asymptotic consistency in D-dimensional search space