NeurGO: Learning to Generate Elite Candidates for Meta-Black-Box Expensive Optimization

NeurGO presents a generative Meta-Black-Box Optimization framework that synthesizes elite candidates directly from historical population states. An attention-based encoder represents population-level search trends, while a conditional decoder generates new solutions without evaluating a large offspring pool. The method adds a quality-diversity loss intended to preserve both solution quality and population diversity. According to the abstract, NeurGO performs better under equal evaluation budgets and converges faster on the CEC 2008 and COCO BBOB suites. However, the abstract does not provide numerical gains, ablations, or details about computational overhead.