Generation of Continuous Optimization Benchmark Problems Guided by Deep Exploratory Landscape Analysis

The development and evaluation of optimization algorithms critically rely on representative benchmark functions. However, existing benchmarks often offer limited coverage of the diverse landscape characteristics encountered in real-world problems. Moreover, real-world black-box optimization tasks involve costly or time-consuming evaluations, highlighting the need for efficient surrogate models. This work presents a novel, feature-driven method for the automated and targeted generation of synthetic benchmark functions using Deep Exploratory Landscape Analysis (Deep ELA) features. The generated functions are neural network models tailored such that their Deep ELA feature vectors closely match user-specified targets. Unlike previous approaches that depend on handcrafted ELA features and computationally expensive surrogate modeling, our method leverages pretrained transformer-based embeddings and gradient-based optimization to directly produce functions exhibiting desired feature combinations. Our method can generate functions whose Deep ELA vectors resemble those of BBOB problems across multiple dimensions, posing similar structural properties and yielding comparable algorithmic performances. Furthermore, our approach supports the generation of novel benchmark instances, with structural characteristics whose feature representations extend beyond those covered by existing test suites. Overall, our proposed approach offers a scalable methodology for creating efficient surrogate models of existing benchmark libraries or expensive real-world problems, and for generating novel optimization problems with much more diverse landscape characteristics.