A functional interface to the CMA-ES optimizer, a (stochastic) non-convex function minimizer
param = cma_optim() param = cma_optim([]) [xopt, f, out, param] = cma_optim(costf, x0, sigma0 [, param])
objective function (cost function) to be minimized. costf must accept a column vector as input and return the scalar cost to be minimized. The return value %nan is allowed and leads to an immediate resampling (and (re-)evaluation) of the solution during the same iteration.
initial solution, a column vector of real values.
the initial standard deviation acting on x0, typically 1/3 of the typical variable ranges. If the ranges of variables are different, use param.scaling_of_variables
optional parameters collected in a struct. '[]' or '{}' invokes all default settings. param.verb controls verbosity. After the objective function is debugged and when the returned solution(s) are satisfactory, logmodulo=0 (no data saving) and displaymodulo=0 (no screen output) make the code completely quiet (and faster).
best found solution vector. On noisy functions and for robust optimization out.solutions.mean.x might be the better choice.
function value of xopt: f=costf(xopt)
additional useful output collected in a struct.
output parameter param contains the actually used parameters (some of them might have been evaluated), including x0 and sigma0.
The function cma_optim uses the functions cma_new() and other cma_*() funcions that implement the CMA-ES in a more object oriented way.
The CMA-ES is a stochastic optimizer. It can make perfect sense to re-run the code several times.
getf('fitfuns.sci'); // define frosen and felli, file should be in CMA-ES-vXXX/demos [f x out] = cma_optim(frosen, rand(8,1), 0.3); // minimize 8-D Rosenbrock function disp(f) // display achieved function value, x is the delivered solution p = cma_optim([]); // get optional parameters disp(p.stop); // lets see the possible and default termination criteria p.readme.stop // lets see what the entries mean p.stop.fitness = 1e-9; // set target function value (minimization) p.verb.plotmodulo = 0; // turn plotting off this time [f x out] = cma_optim(felli, ones(15,1), 1, p); // minimize 15-D hyper-ellipsoid function disp(out.solutions.bestever.f); // show the best-ever function value disp(felli(out.solutions.mean.x)); // might be an even better solution cma_plot; // let's finally plot the run | ![]() | ![]() |
cma_new()
cma_ask()
cma_tell()
...