Enhancing Mutation Matrix Adaptation Evolution Strategy with Orthogonal Sampling

This paper explores orthogonal sampling to enhance Evolution Strategies, focusing on mutation matrix adaptation (MMA-ES), a variant of CMA-ES. We propose using QR decomposition to generate random orthogonal search directions, overcoming numerical instability associated with Gram-Schmidt orthogonalization. The resulting approach, termed oMMA-ES, demonstrates remarkable performance on standard test functions. Surprisingly, our experimental findings indicate that pairwise selection, commonly used to accelerate convergence, may hinder algorithm progress. Orthogonal sampling with QR decomposition presents a promising strategy to boost the efficiency and effectiveness of MMA-ES for optimization tasks.