Title: CoAction: Cross-task Correlation-aware Pareto Set Learning

Abstract: Pareto set learning (PSL) has emerged as a significant approach in multi-objective optimization, utilizing neural networks to map preference vectors directly to Pareto optimal solutions. Despite its promise, current PSL methodologies are largely restricted to addressing individual multi-objective problems in isolation. This narrow scope presents two major drawbacks: it inflates computational expenses in multitask optimization contexts by necessitating distinct models for every task, and it overlooks the valuable inter-task correlations that exist across different problems. To overcome these challenges, we introduce CoAction (Cross-tAsk correlation-aware Pareto Set Learning), a novel framework designed to manage multiple tasks concurrently through the use of a task-aware transformer. By incorporating task-specific embedding vectors for each task, the model not only accurately differentiates between tasks but also promotes effective knowledge transfer among them. The core architecture employs a Transformer encoder, leveraging its self-attention mechanism to identify and utilize complex dependencies between tasks. Our comprehensive evaluation, conducted on extensive multitask test suites that include both benchmark problems and real-world applications, confirms the efficacy of the proposed method, which delivers competitive results in terms of Hypervolume, Range, and Sparsity metrics.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC