Safe ASNG method enables risk-free binary optimization from Uchida et al.

Researchers Kento Uchida, Ryoki Hamano, Masahiro Nomura, and Shinichi Shirakawa have developed Safe ASNG, a novel algorithm for safe optimization on binary search spaces. The method extends the adaptive stochastic natural gradient method (ASNG) by incorporating safety constraints that keep a safety function non-negative during optimization. It constructs surrogate models using discrete Walsh functions to estimate Lipschitz constants with respect to Hamming distance, then computes a safe region around previously evaluated safe solutions. New candidate solutions are projected to their nearest neighbors within that region, effectively suppressing unsafe evaluations. In benchmark experiments on binary domains, Safe ASNG achieved efficient optimization while failing to evaluate unsafe solutions, unlike comparative methods that did not maintain safety.

The work addresses a critical gap: while safe optimization methods exist for continuous spaces, binary spaces have been largely overlooked despite their prevalence in medical diagnostics, engineering design, and combinatorial optimization. The use of Walsh functions is mathematically elegant because they form a complete orthonormal basis on binary hypercubes, enabling accurate Lipschitz constant estimation. This allows the algorithm to adaptively balance exploration and safety. The paper has been accepted as a full paper at GECCO2026, the leading conference on genetic and evolutionary computation. For practitioners, Safe ASNG offers a practical tool for optimizing binary parameters (e.g., feature selection, logic circuit design, drug molecule binary fingerprints) where unsafe candidates could lead to costly or dangerous outcomes.