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

In expensive black-box optimization (BBO), function evaluations are extremely costly, making traditional evolutionary algorithms and Bayesian optimization inefficient due to their reliance on large offspring pools or surrogate models that often fail under tight budgets. This paper asks whether a generative meta-learning approach can learn to synthesize elite candidate solutions directly from historical population data, thereby bypassing the need for expensive surrogate construction or large-scale candidate evaluation.

The authors propose NeurGO, an end-to-end generative framework consisting of two primary components: a Population Context Encoder (PCE) and an Elite Synthesis Decoder (ESD). The PCE uses a Transformer-based architecture with rank-fitness dual-aware attention to capture global search trends and evolutionary context from the current population. The ESD then takes these refined features and synthesizes a small, high-quality batch of candidate solutions. To ensure the model balances exploration and exploitation, the authors train the framework using a Quality-Diversity (QD) loss, which maximizes the improvement of the best candidate while simultaneously penalizing candidate homogeneity.

NeurGO was evaluated on the CEC 2008 and COCO BBOB benchmark suites. The results demonstrate that NeurGO consistently achieves faster convergence and superior final optimization performance compared to traditional evolutionary algorithms, surrogate-assisted methods, and existing MetaBBO baselines. By directly synthesizing candidates rather than relying on surrogate-guided search, the model effectively minimizes wasted evaluations, making it highly suitable for scenarios where the evaluation budget is severely constrained.