HEMO evolves the search harness for multi-objective optimization
Original title: HEMO: Harness Self-Evolution for Multi-Objective Combinatorial Optimization
The method updates one harness module at a time using solution quality, runtime, and a discounted UCB scheduler.
HEMO treats the automatic operator-design harness, not only the operators, as a self-evolving target for multi-objective combinatorial optimization. It keeps a persistent harness state and revises one module at a time, covering experience retrieval, operator improvement, parent selection, and search orchestration. Downstream solution quality and runtime guide updates, while a discounted UCB scheduler shifts effort toward promising modules. A persistent memory filesystem carries candidate outcomes and diagnostic feedback across states. On eight MOCO tasks and two evolutionary backbones, the authors report stronger results than conventional evolutionary, neural, and LLM-based AOD baselines, plus transfer to related tasks.
Unconfirmed: Performance and transfer claims are the authors' experimental results and are not independently confirmed.
- Most AOD methods evolve operators while leaving the discovery process fixed.
- HEMO revises retrieval, improvement, parent selection, and orchestration logic.
- A discounted UCB scheduler directs adaptation as module utility changes.
- Experiments cover eight MOCO tasks and two evolutionary backbones.
- Authors report transfer of the evolved harness to related tasks.
From the original article
1 min read at arxiv.orgAutomatic operator design (AOD) has increasingly adopted large language models (LLMs) to generate problem-specific search and variation operators for multi-objective combinatorial optimization (MOCO). Yet most methods evolve the operators while leaving the process that discovers them fixed. This limits adaptation as new Pareto trade-offs and operator interactions change where and how the search should proceed.
By Jianghan Zhu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.