Choosing an Algorithm

Which algorithm should I use?

Choose based on the number of parameters and the nature of your objective function.

Algorithm	Suitable for	Characteristics
minsearch	Few parameters (~10), quick local minimum search	Nelder-Mead method. Fast but may get trapped in local minima. Gradient-free
global_search	Global optimization of continuous parameters	Global optimizers from scipy.optimize (differential evolution, SHGO, DIRECT, dual annealing). Requires scipy. Supports MPI-parallel candidate evaluation
bayes	Expensive objective function, minimize evaluation count	Surrogate model via Gaussian process regression. Minimizes evaluations. Requires physbo
mapper	Overview of parameter space. Few parameters (2-3)	Evaluates all grid points. Computational cost grows explosively with the number of parameters
exchange	Multimodal objective function, broad search	Replica exchange method. Avoids local minima. MPI parallelism recommended (requires mpi4py; single-process runs also work)
pamc	Posterior distribution estimation, model evidence calculation	Population annealing. Suitable for statistical estimation
random_search	Quick overview of parameter space	Random sampling. Simple and robust
ttopt	Finding minima in problems with relatively many parameters or with discrete variables	Tensor train optimization. Gradient-free cross approximation. Supports MPI