def reuse_cma_run():
# problem definition
d = 20
f = f_cont.Rosenbrock(d, target_eval=1e-10, max_eval=1e6)
init_m, init_sigma = f_cont.initial_setting_for_gaussian(f)
# parameter setting
lam = cma.CMAParam.pop_size(d)
K = 5
# weight function
non_inc_f = weight.CMANonIncFunc()
w_func = weight.QuantileBasedWeight(non_inc_f,
tie_case=True,
normalization=False,
min_problem=True)
# optimizer
opt = reuse_cma.ReuseCMA(d,
w_func,
lam,
K=K,
m=init_m,
sigma=init_sigma,
csa=True,
rank1=True,
reuse_m=False)
# run
return opt.run(sampler.DefaultSampler(f, lam), logger=None, verbose=True)
def gaussian_reuse_igo():
# problem definition
d = 20
f = f_cont.Rosenbrock(d, target_eval=1e-10, max_eval=1e6)
init_m, init_sigma = f_cont.initial_setting_for_gaussian(f)
# parameter setting
lam = cma.CMAParam.pop_size(d)
K = 5
# weight function
non_inc_f = weight.CMANonIncFunc()
w_func = weight.QuantileBasedWeight(non_inc_f,
tie_case=True,
normalization=False,
min_problem=True)
# learning rate
eta_c = cma.CMAParam.c_mu(d, cma.CMAParam.mu_eff(lam))
# optimizer
opt = reuse_igo.ReuseGaussianIGO(d,
w_func,
lam,
K=K,
m=init_m,
C=(init_sigma**2) * np.identity(d),
eta_m=1.,
eta_C=eta_c)
# run
return opt.run(sampler.DefaultSampler(f, lam), logger=None, verbose=True)
def bernoulli_reuse_igo():
# problem definition
d = 32
f = f_bin.OneMax(d, max_eval=d**2)
# f = f_bin.LeadingOne(dim, max_eval=dim**3)
# parameter setting
lam = 2
K = 5
theta_min = 1. / d
theta_max = 1. - 1. / d
# weight function
non_inc_f = weight.SelectionNonIncFunc(threshold=0.25,
negative_weight=True)
w = weight.QuantileBasedWeight(non_inc_f=non_inc_f,
tie_case=True,
normalization=False,
min_problem=False)
# learning rate
eta = 1. / d
# model
opt = reuse_igo.ReuseBernoulliIGO(d,
w,
lam,
K=K,
eta=eta,
theta_max=theta_max,
theta_min=theta_min)
# run
return opt.run(sampler.DefaultSampler(f, lam), logger=None, verbose=True)
def cma_run():
# problem definition
d = 20
f = f_cont.Rosenbrock(d, target_eval=1e-10, max_eval=1e6)
init_m, init_sigma = f_cont.initial_setting_for_gaussian(f)
lam = cma.CMAParam.pop_size(d)
# weight function
w_func = weight.CMAWeight(lam)
# optimizer
opt = cma.CMAES(d, w_func, m=init_m, sigma=init_sigma)
# run
return opt.run(sampler.DefaultSampler(f, lam), logger=None, verbose=True)
def gaussian_igo():
# problem definition
d = 20
f = f_cont.Ellipsoid(d, target_eval=1e-10, max_eval=1e6)
init_m, init_sigma = f_cont.initial_setting_for_gaussian(f)
# parameter setting
lam = cma.CMAParam.pop_size(d)
# weight function
w_func = weight.CMAWeight(lam)
# learning rate
eta_c = cma.CMAParam.c_mu(d, cma.CMAParam.mu_eff(lam))
# optimizer
opt = igo.GaussianIGO(d,
w_func,
m=init_m,
C=(init_sigma**2) * np.identity(d),
eta_m=1.0,
eta_C=eta_c)
# run
return opt.run(sampler.DefaultSampler(f, lam), logger=None, verbose=True)