Scenario-based self-learning transfer framework for multi-task optimization problems

Zhuoming Yuan

doi:10.1016/j.knosys.2025.113824

Scenario-based self-learning transfer framework for multi-task optimization problems

2025-06-06

0

PRE

AI

DOI：10.1016/j.knosys.2025.113824

Original

Original request for help

Save

Abstract

En 中文

Evolutionary multi-task optimization (EMTO), which tackles multi-task optimization problems (MTOPs) simultaneously, is gaining popularity. Yet, there are still two problems that have not been substantively addressed: (1) how to design the strategy to efficiently solve MTOPs in a variety of evolutionary scenarios; (2) how to automatically adjust the designed strategy when solving MTOPs with a variety of evolutionary scenarios. Herein, we find that learning the relationship mapping between the evolutionary scenario and the scenario-specific strategy plays a crucial role for the issues raised, and therefore propose a novel scenario-based self-learning transfer (SSLT) framework. In SSLT, for the first issue, we categorize the scenarios into four possible situations in the MTOP environment, thereby designing a set of corresponding scenario-specific strategies, and we also design an ensemble method to characterize the scenario in terms of intra-task and inter-task scenario features. For the second issue, SSLT adopts the deep Q-network as the relationship mapping model to learn the relationship mapping between the evolutionary scenario and the scenario-specific strategy. We conducted experiments to compare SSLT-based algorithms with state-of-the-art competitors on two sets of MTOPs and real-world interplanetary trajectory design missions. The experimental results confirm the favorable performance of the SSLT-based algorithms against competitors.

Scenario-based self-learning transfer framework for multi-task optimization problems

Abstract

Journal

Organization