“…42 In ML, it is common to simply use various random initialisations and local minimisation, which has limited use in physical systems. 43 Examples of random, or pseudo-random, initialisation and minimisation are seen in K -means, 44–46 Gaussian processes, 47 and neural networks. All these methods rely upon minimisation algorithms to locate local minima of the cost function surfaces, and common choices are limited-memory 48,49 quasi-Newton Broyden, 50 Fletcher, 51 Goldfarb, 52 Shanno 53 (L-BFGS) routine and its variants with box-constraints 54,55 for bounded problems such as fitting Gaussian processes.…”