This paper introduces a full invariance oriented evolution strategies algorithm, SYNCMA, which effectively competes with leading Bayesian optimization methods in high-dimensional tasks. The researchers first build the framework INVIGO that incorporates historical information while maintaining invariance and computational complexity. They then apply INVIGO on multi-dimensional Gaussian, resulting in an invariant and scalable optimizer, SYNCMA. The paper shows SYNCMA’s advantages over other Gaussian-based evolution strategies and benchmarks it against leading algorithms in Bayesian optimization on various high-dimension tasks. SYNCMA demonstrates great competence, if not dominance, over other algorithms in sample efficiency.
Publication date: 3 Jan 2024
Project Page: https://arxiv.org/abs/2401.01579v1
Paper: https://arxiv.org/pdf/2401.01579
