Standard noise test scenario

[nikohansen]

import functools
import cma

fun = functools.partial(cma.ff.noisysphere, cond=1e4, noise=1, noise_offset=0.0)
x, es = cma.fmin2(fun, 33 * [1], 1, {
                    'AdaptSigma': cma.sigma_adaptation.CMAAdaptSigmaTPA,
                    'ftarget': 1e-8, 'verb_disp_overwrite': 600
                                     },
                  noise_handler=True  # means cma.NoiseHandler
                 )
es.plot()

works only with noise handler (or a different step-size bias).

[nikohansen]

Another usecase where not only sigma-increase but also reevaluations are needed:

import cma

def fun(x):
    return cma.ff.noisysphere(x, cond=1e4, noise_offset=0, noise=8.1)

x0 = 14 * [-0.1]
sigma0 = 1
xopt, es = cma.fmin2(fun, x0, sigma0, {'verb_disp_overwrite':400, 'maxfevals': 5e6,
        },
        noise_handler=cma.NoiseHandler(len(x0), [1, 1, 100]))
es.plot();