Context-Aware Contrastive Surrogates for Scaling-up Expensive Multiobjective Optimization

ICLR 2026 Conference Submission18040 Authors

Keywords: Expensive Multiobjective Optimization, Surrogate-assisted Evolutionary Algorithm, Representation Learning

TL;DR: FSMOEA captures population-level context to improve surrogate prediction accuracy, leverages a low-dimensional latent space to accelerate evolutionary search, and employs lightweight architectures to reduce computational overhead.

Abstract: Addressing expensive multiobjective optimization problems (EMOPs) poses a significant challenge due to the high cost of objective evaluations. We propose FSMOEA, a scalable and efficient framework that enhances surrogate-assisted multiobjective evolutionary algorithms (SMOEAs) by introducing foresighted surrogate models. FSMOEA captures population-level context to improve surrogate prediction accuracy, leverages a low-dimensional latent space to accelerate evolutionary search, and employs lightweight models to reduce computational overhead. Designed for plug-and-play integration, the foresight model can be embedded into existing contrastive (i.e., classification- and relation-based) SMOEAs, improving performance on scaling-up EMOPs. We provide theoretical analysis that formalizes the benefits of population-aware representation and latent-space optimization. Extensive experiments on 107 benchmarks show that FSMOEA consistently outperforms state-of-the-art methods in both convergence speed and optimization quality. Source code is attached and will be available at Linkxxx.

Primary Area: optimization

Submission Number: 18040