A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization

doi:10.1109/tevc.2026.3727397

A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization

2026-08-26

0

PRE

AI

DOI：10.1109/tevc.2026.3727397

Original

Original request for help

Save

Abstract

En 中文

Dynamic multi-objective optimization problems require algorithms to track the time-varying Pareto set and Pareto front as environments change. This is particularly difficult under irregular changes and sparse observations, which weaken short-term prediction, increase the risk of negative transfer, and aggravate the cold-start problem in online adaptation. We therefore propose DB-GEN, a decoupled basis-vector-driven generative framework for online population reinitialization. DB-GEN uses frequency decoupling based on the discrete wavelet transform to represent evolutionary trajectories across multiple scales and sparse dictionary learning to extract transferable basis vectors. A topology-aware contrastive loss then guides their recombination into a structured latent manifold, from which surrogate-assisted generative search constructs initial populations for new environments. Pre-trained offline on 120 million solutions, DB-GEN can directly handle unseen target problems whose dynamics can be represented by learned basis combinations, without target-specific retraining or fine-tuning. Across 76 configurations, DB-GEN achieves the best MIGD in 57 cases and the second-best in nine.

Keywords:

Dynamic multi-objective optimization

evolutionary algorithms

structure learning

latent manifold

A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization

Abstract

Journal

Organization

Cited Papers