Low-Memory Matrix Adaptation Evolution Strategies Exploiting Gradient Information and Lévy Flight

The Low-Memory Matrix Adaptation Evolution Strategy is a recent variant of CMA-ES that is specifically meant for large-scale numerical optimization. In this paper, we investigate if and how gradient information can be included in this algorithm, in order to enhance its performance. Furthermore, we consider the incorporation of Lévy flight to alleviate stability issues due to possibly unreliably gradient estimation as well as promote better exploration. In total, we propose four new variants of LMMA-ES, making use of real and estimated gradient, with and without Lévy flight. We test the proposed variants on two neural network training tasks, one for image classification through the newly introduced Forward-Forward paradigm, and one for a Reinforcement Learning problem, as well as five benchmark functions for numerical optimization.