NeurGO: Generative Framework for Elite Candidate Synthesis in Expensive Optimization
A new framework named NeurGO has been introduced by researchers, designed as a generative Meta-Black-Box Optimization (MetaBBO) system that synthesizes top candidates from past population states, thereby eliminating unnecessary evaluations of subpar solutions. Conventional evolutionary algorithms and MetaBBO methods tend to waste evaluations during candidate selection, while surrogate-assisted techniques often face challenges related to model bias and local optima. NeurGO employs an attention-based encoder to identify population-wide search patterns and utilizes a decoder to produce high-quality candidates. This framework effectively tackles costly black-box optimization challenges prevalent in scientific and engineering fields, where function evaluations are expensive and budgets are constrained. The findings are available on arXiv with the identifier 2607.23408.
Key facts
- NeurGO is a generative MetaBBO framework.
- It synthesizes elite candidates from historical population states.
- Uses an attention-based encoder to capture search trends.
- Decoder generates high-quality candidates conditioned on encoder representation.
- Aims to reduce wasted evaluations on inferior solutions.
- Addresses expensive black-box optimization with limited budgets.
- Published on arXiv with ID 2607.23408.
- Announce type is new.
Entities
Institutions
- arXiv