def result(self):
""" The OptimizeResult """
res = OptimizeResult()
res.x = self._xmin
res.fun = self._fvalue
res.message = self._message
res.nit = self._step_record
return res
def result(self):
""" The OptimizeResult """
res = OptimizeResult()
res.x = self._xmin
res.fun = self._fvalue
res.message = self._message
res.nit = self._step_record
return res
def get_optimization_results(t, N, factors, taskset):
K = len(factors)
results = []
for k in range(K):
factor = factors[k]
result = OptimizeResult()
result.x = deepcopy(factor.theta)
result.fun = factor.f_opt
result.nit = t
result.nfev = (t + 1) * 2 * N
result.message = deepcopy(taskset.normalizers)
results.append(result)
return results
def _int_spontaneous_raman(self, z_array, raman_matrix, alphap_fiber,
freq_array, cr_raman_matrix, freq_diff, ase_bc,
bn_array, temperature):
spontaneous_raman_scattering = OptimizeResult()
dx = self.solver_params.z_resolution
h = ph.value('Planck constant')
kb = ph.value('Boltzmann constant')
power_ase = np.nan * np.ones(raman_matrix.shape)
int_pump = cumtrapz(raman_matrix, z_array, dx=dx, axis=1, initial=0)
for f_ind, f_ase in enumerate(freq_array):
cr_raman = cr_raman_matrix[f_ind, :]
vibrational_loss = f_ase / freq_array[:f_ind]
eta = 1 / (np.exp(
(h * freq_diff[f_ind, f_ind + 1:]) / (kb * temperature)) - 1)
int_fiber_loss = -alphap_fiber[f_ind] * z_array
int_raman_loss = np.sum(
(cr_raman[:f_ind] * vibrational_loss *
int_pump[:f_ind, :].transpose()).transpose(),
axis=0)
int_raman_gain = np.sum(
(cr_raman[f_ind + 1:] *
int_pump[f_ind + 1:, :].transpose()).transpose(),
axis=0)
int_gain_loss = int_fiber_loss + int_raman_gain + int_raman_loss
new_ase = np.sum(
(cr_raman[f_ind + 1:] * (1 + eta) *
raman_matrix[f_ind + 1:, :].transpose()).transpose() * h *
f_ase * bn_array[f_ind],
axis=0)
bc_evolution = ase_bc[f_ind] * np.exp(int_gain_loss)
ase_evolution = np.exp(int_gain_loss) * cumtrapz(
new_ase * np.exp(-int_gain_loss), z_array, dx=dx, initial=0)
power_ase[f_ind, :] = bc_evolution + ase_evolution
spontaneous_raman_scattering.x = power_ase
spontaneous_raman_scattering.success = True
spontaneous_raman_scattering.message = "Spontaneous Raman Scattering evaluated successfully"
return spontaneous_raman_scattering
def scipy_nlopt_cobyla(*args, **kwargs):
"""Wraps nlopt library cobyla function to be compatible with scipy optimize
parameters:
args[0]: target, function to be minimized
args[1]: x0, starting point for minimization
bounds: list of bounds for the movement
[[min, max], [min, max], ...]
ftol_rel: same as in nlopt
xtol_rel: same as in nlopt
one of the tol_rel should be specified
returns:
OptimizeResult() object with properly set x, fun, success.
status is not set when nlopt.RoundoffLimited is raised
"""
answ = OptimizeResult()
bounds = kwargs['bounds']
opt = nlopt.opt(nlopt.LN_COBYLA, len(args[1]))
opt.set_lower_bounds([i[0] for i in bounds])
opt.set_upper_bounds([i[1] for i in bounds])
if 'ftol_rel' in kwargs.keys():
opt.set_ftol_rel(kwargs['ftol_rel'])
if 'xtol_rel' in kwargs.keys():
opt.set_ftol_rel(kwargs['xtol_rel'])
opt.set_min_objective(args[0])
x0 = list(args[1])
try:
x1 = opt.optimize(x0)
except nlopt.RoundoffLimited:
answ.x = x0
answ.fun = args[0](x0)
answ.success = False
answ.message = 'nlopt.RoundoffLimited'
return answ
answ.x = x1
answ.fun = args[0](x1)
answ.success = True if opt.last_optimize_result() in [3, 4] else False
answ.status = opt.last_optimize_result()
if not answ.fun == opt.last_optimum_value():
print 'Something\'s wrong, ', answ.fun, opt.last_optimum_value()
return answ
def get_optimization_results(t, population, factorial_cost, scalar_fitness,
skill_factor, message):
K = len(set(skill_factor))
N = len(population) // 2
results = []
for k in range(K):
result = OptimizeResult()
x, fun = get_best_individual(population, factorial_cost,
scalar_fitness, skill_factor, k)
result.x = x
result.fun = fun
result.message = message
result.nit = t
result.nfev = (t + 1) * N
mean, std = get_statistics(factorial_cost, skill_factor, k)
result.mean = mean
result.std = std
results.append(result)
return results
def dual_annealing(func,
bounds,
args=(),
maxiter=1000,
local_search_options={},
initial_temp=5230.,
restart_temp_ratio=2.e-5,
visit=2.62,
accept=-5.0,
maxfun=1e7,
seed=None,
no_local_search=False,
callback=None,
x0=None):
"""
Find the global minimum of a function using Dual Annealing.
Parameters
----------
func : callable
The objective function to be minimized. Must be in the form
``f(x, *args)``, where ``x`` is the argument in the form of a 1-D array
and ``args`` is a tuple of any additional fixed parameters needed to
completely specify the function.
bounds : sequence, shape (n, 2)
Bounds for variables. ``(min, max)`` pairs for each element in ``x``,
defining bounds for the objective function parameter.
args : tuple, optional
Any additional fixed parameters needed to completely specify the
objective function.
maxiter : int, optional
The maximum number of global search iterations. Default value is 1000.
local_search_options : dict, optional
Extra keyword arguments to be passed to the local minimizer
(`minimize`). Some important options could be:
``method`` for the minimizer method to use and ``args`` for
objective function additional arguments.
initial_temp : float, optional
The initial temperature, use higher values to facilitates a wider
search of the energy landscape, allowing dual_annealing to escape
local minima that it is trapped in. Default value is 5230. Range is
(0.01, 5.e4].
restart_temp_ratio : float, optional
During the annealing process, temperature is decreasing, when it
reaches ``initial_temp * restart_temp_ratio``, the reannealing process
is triggered. Default value of the ratio is 2e-5. Range is (0, 1).
visit : float, optional
Parameter for visiting distribution. Default value is 2.62. Higher
values give the visiting distribution a heavier tail, this makes
the algorithm jump to a more distant region. The value range is (0, 3].
accept : float, optional
Parameter for acceptance distribution. It is used to control the
probability of acceptance. The lower the acceptance parameter, the
smaller the probability of acceptance. Default value is -5.0 with
a range (-1e4, -5].
maxfun : int, optional
Soft limit for the number of objective function calls. If the
algorithm is in the middle of a local search, this number will be
exceeded, the algorithm will stop just after the local search is
done. Default value is 1e7.
seed : {int or `~numpy.random.mtrand.RandomState` instance}, optional
If `seed` is not specified the `~numpy.random.mtrand.RandomState`
singleton is used.
If `seed` is an int, a new ``RandomState`` instance is used,
seeded with `seed`.
If `seed` is already a ``RandomState`` instance, then that
instance is used.
Specify `seed` for repeatable minimizations. The random numbers
generated with this seed only affect the visiting distribution
function and new coordinates generation.
no_local_search : bool, optional
If `no_local_search` is set to True, a traditional Generalized
Simulated Annealing will be performed with no local search
strategy applied.
callback : callable, optional
A callback function with signature ``callback(x, f, context)``,
which will be called for all minima found.
``x`` and ``f`` are the coordinates and function value of the
latest minimum found, and ``context`` has value in [0, 1, 2], with the
following meaning:
- 0: minimum detected in the annealing process.
- 1: detection occurred in the local search process.
- 2: detection done in the dual annealing process.
If the callback implementation returns True, the algorithm will stop.
x0 : ndarray, shape(n,), optional
Coordinates of a single N-D starting point.
Returns
-------
res : OptimizeResult
The optimization result represented as a `OptimizeResult` object.
Important attributes are: ``x`` the solution array, ``fun`` the value
of the function at the solution, and ``message`` which describes the
cause of the termination.
See `OptimizeResult` for a description of other attributes.
Notes
-----
This function implements the Dual Annealing optimization. This stochastic
approach derived from [3]_ combines the generalization of CSA (Classical
Simulated Annealing) and FSA (Fast Simulated Annealing) [1]_ [2]_ coupled
to a strategy for applying a local search on accepted locations [4]_.
An alternative implementation of this same algorithm is described in [5]_
and benchmarks are presented in [6]_. This approach introduces an advanced
method to refine the solution found by the generalized annealing
process. This algorithm uses a distorted Cauchy-Lorentz visiting
distribution, with its shape controlled by the parameter :math:`q_{v}`
.. math::
g_{q_{v}}(\\Delta x(t)) \\propto \\frac{ \\
\\left[T_{q_{v}}(t) \\right]^{-\\frac{D}{3-q_{v}}}}{ \\
\\left[{1+(q_{v}-1)\\frac{(\\Delta x(t))^{2}} { \\
\\left[T_{q_{v}}(t)\\right]^{\\frac{2}{3-q_{v}}}}}\\right]^{ \\
\\frac{1}{q_{v}-1}+\\frac{D-1}{2}}}
Where :math:`t` is the artificial time. This visiting distribution is used
to generate a trial jump distance :math:`\\Delta x(t)` of variable
:math:`x(t)` under artificial temperature :math:`T_{q_{v}}(t)`.
From the starting point, after calling the visiting distribution
function, the acceptance probability is computed as follows:
.. math::
p_{q_{a}} = \\min{\\{1,\\left[1-(1-q_{a}) \\beta \\Delta E \\right]^{ \\
\\frac{1}{1-q_{a}}}\\}}
Where :math:`q_{a}` is a acceptance parameter. For :math:`q_{a}<1`, zero
acceptance probability is assigned to the cases where
.. math::
[1-(1-q_{a}) \\beta \\Delta E] < 0
The artificial temperature :math:`T_{q_{v}}(t)` is decreased according to
.. math::
T_{q_{v}}(t) = T_{q_{v}}(1) \\frac{2^{q_{v}-1}-1}{\\left( \\
1 + t\\right)^{q_{v}-1}-1}
Where :math:`q_{v}` is the visiting parameter.
.. versionadded:: 1.2.0
References
----------
.. [1] Tsallis C. Possible generalization of Boltzmann-Gibbs
statistics. Journal of Statistical Physics, 52, 479-487 (1998).
.. [2] Tsallis C, Stariolo DA. Generalized Simulated Annealing.
Physica A, 233, 395-406 (1996).
.. [3] Xiang Y, Sun DY, Fan W, Gong XG. Generalized Simulated
Annealing Algorithm and Its Application to the Thomson Model.
Physics Letters A, 233, 216-220 (1997).
.. [4] Xiang Y, Gong XG. Efficiency of Generalized Simulated
Annealing. Physical Review E, 62, 4473 (2000).
.. [5] Xiang Y, Gubian S, Suomela B, Hoeng J. Generalized
Simulated Annealing for Efficient Global Optimization: the GenSA
Package for R. The R Journal, Volume 5/1 (2013).
.. [6] Mullen, K. Continuous Global Optimization in R. Journal of
Statistical Software, 60(6), 1 - 45, (2014). DOI:10.18637/jss.v060.i06
Examples
--------
The following example is a 10-D problem, with many local minima.
The function involved is called Rastrigin
(https://en.wikipedia.org/wiki/Rastrigin_function)
>>> from scipy.optimize import dual_annealing
>>> func = lambda x: np.sum(x*x - 10*np.cos(2*np.pi*x)) + 10*np.size(x)
>>> lw = [-5.12] * 10
>>> up = [5.12] * 10
>>> ret = dual_annealing(func, bounds=list(zip(lw, up)), seed=1234)
>>> print("global minimum: xmin = {0}, f(xmin) = {1:.6f}".format(
... ret.x, ret.fun))
global minimum: xmin = [-4.26437714e-09 -3.91699361e-09 -1.86149218e-09 -3.97165720e-09
-6.29151648e-09 -6.53145322e-09 -3.93616815e-09 -6.55623025e-09
-6.05775280e-09 -5.00668935e-09], f(xmin) = 0.000000
""" # noqa: E501
if x0 is not None and not len(x0) == len(bounds):
raise ValueError('Bounds size does not match x0')
lu = list(zip(*bounds))
lower = np.array(lu[0])
upper = np.array(lu[1])
# Check that restart temperature ratio is correct
if restart_temp_ratio <= 0. or restart_temp_ratio >= 1.:
raise ValueError('Restart temperature ratio has to be in range (0, 1)')
# Checking bounds are valid
if (np.any(np.isinf(lower)) or np.any(np.isinf(upper))
or np.any(np.isnan(lower)) or np.any(np.isnan(upper))):
raise ValueError('Some bounds values are inf values or nan values')
# Checking that bounds are consistent
if not np.all(lower < upper):
raise ValueError('Bounds are not consistent min < max')
# Checking that bounds are the same length
if not len(lower) == len(upper):
raise ValueError('Bounds do not have the same dimensions')
# Wrapper for the objective function
func_wrapper = ObjectiveFunWrapper(func, maxfun, *args)
# Wrapper fot the minimizer
minimizer_wrapper = LocalSearchWrapper(bounds, func_wrapper,
**local_search_options)
# Initialization of RandomState for reproducible runs if seed provided
rand_state = check_random_state(seed)
# Initialization of the energy state
energy_state = EnergyState(lower, upper, callback)
energy_state.reset(func_wrapper, rand_state, x0)
# Minimum value of annealing temperature reached to perform
# re-annealing
temperature_restart = initial_temp * restart_temp_ratio
# VisitingDistribution instance
visit_dist = VisitingDistribution(lower, upper, visit, rand_state)
# Strategy chain instance
strategy_chain = StrategyChain(accept, visit_dist, func_wrapper,
minimizer_wrapper, rand_state, energy_state)
need_to_stop = False
iteration = 0
message = []
# OptimizeResult object to be returned
optimize_res = OptimizeResult()
optimize_res.success = True
optimize_res.status = 0
t1 = np.exp((visit - 1) * np.log(2.0)) - 1.0
# Run the search loop
while (not need_to_stop):
for i in range(maxiter):
# Compute temperature for this step
s = float(i) + 2.0
t2 = np.exp((visit - 1) * np.log(s)) - 1.0
temperature = initial_temp * t1 / t2
if iteration >= maxiter:
message.append("Maximum number of iteration reached")
need_to_stop = True
break
# Need a re-annealing process?
if temperature < temperature_restart:
energy_state.reset(func_wrapper, rand_state)
break
# starting strategy chain
val = strategy_chain.run(i, temperature)
if val is not None:
message.append(val)
need_to_stop = True
optimize_res.success = False
break
# Possible local search at the end of the strategy chain
if not no_local_search:
val = strategy_chain.local_search()
if val is not None:
message.append(val)
need_to_stop = True
optimize_res.success = False
break
iteration += 1
# Setting the OptimizeResult values
optimize_res.x = energy_state.xbest
optimize_res.fun = energy_state.ebest
optimize_res.nit = iteration
optimize_res.nfev = func_wrapper.nfev
optimize_res.njev = func_wrapper.ngev
optimize_res.nhev = func_wrapper.nhev
optimize_res.message = message
return optimize_res
def model_policy_gradient(
f: Callable[..., float],
x0: np.ndarray,
*,
args=(),
learning_rate: float = 1e-2,
decay_rate: float = 0.96,
decay_steps: int = 5,
log_sigma_init: float = -5.0,
max_iterations: int = 1000,
batch_size: int = 10,
radius_coeff: float = 3.0,
warmup_steps: int = 10,
batch_size_model: int = 65536,
save_func_vals: bool = False,
random_state: "cirq.RANDOM_STATE_OR_SEED_LIKE" = None,
known_values: Optional[Tuple[List[np.ndarray], List[float]]] = None,
max_evaluations: Optional[int] = None
) -> scipy.optimize.OptimizeResult:
"""Model policy gradient algorithm for black-box optimization.
The idea of this algorithm is to perform policy gradient, but estimate
the function values using a surrogate model.
The surrogate model is a least-squared quadratic
fit to points sampled from the vicinity of the current iterate.
Args:
f: The function to minimize.
x0: An initial guess.
args: Additional arguments to pass to the function.
learning_rate: The learning rate for the policy gradient.
decay_rate: the learning decay rate for the Adam optimizer.
decay_steps: the learning decay steps for the Adam optimizer.
log_sigma_init: the intial value for the sigma of the policy
in the log scale.
max_iterations: The maximum number of iterations to allow before
termination.
batch_size: The number of points to sample in each iteration. The cost
of evaluation of these samples are computed through the
quantum computer cost model.
radius_coeff: The ratio determining the size of the radius around
the current iterate to sample points from to build the quadratic model.
The ratio is with respect to the maximal ratio of the samples
from the current policy.
warmup_steps: The number of steps before the model policy gradient is performed.
before these steps, we use the policy gradient without the model.
batch_size_model: The model sample batch size.
After we fit the quadratic model, we use the model to evaluate
on big enough batch of samples.
save_func_vals: whether to compute and save the function values for
the current value of parameter.
random_state: A seed (int) or `np.random.RandomState` class to use when
generating random values. If not set, defaults to using the module
methods in `np.random`.
known_values: Any prior known values of the objective function.
This is given as a tuple where the first element is a list
of points and the second element is a list of the function values
at those points.
max_evaluations: The maximum number of function evaluations to allow
before termination.
Returns:
Scipy OptimizeResult
"""
random_state = value.parse_random_state(random_state)
if known_values is not None:
known_xs, known_ys = known_values
known_xs = [np.copy(x) for x in known_xs]
known_ys = [np.copy(y) for y in known_ys]
else:
known_xs, known_ys = [], []
if max_evaluations is None:
max_evaluations = np.inf
n = len(x0)
log_sigma = np.ones(n) * log_sigma_init
sigma = np.exp(log_sigma)
# set up the first and second moment estimate
m_mean = np.zeros(n)
v_mean = np.zeros(n)
m_log_sigma = np.zeros(n)
v_log_sigma = np.zeros(n)
# set up lr schedule and optimizer
lr_schedule1 = _ExponentialSchedule(learning_rate,
decay_steps=decay_steps,
decay_rate=decay_rate,
staircase=True)
lr_schedule2 = _ExponentialSchedule(learning_rate,
decay_steps=decay_steps,
decay_rate=decay_rate,
staircase=True)
_, f = wrap_function(f, args)
res = OptimizeResult()
current_x = np.copy(x0)
res.x_iters = [] # initializes as lists
res.xs_iters = []
res.ys_iters = []
res.func_vals = []
res.fun = 0
total_evals = 0
num_iter = 0
message = None
# stats
history_max = -np.inf
while num_iter < max_iterations:
# get samples from the current policy to evaluate
z = random_state.randn(batch_size, n)
new_xs = sigma * z + current_x
if total_evals + batch_size > max_evaluations:
message = "Reached maximum number of evaluations."
break
# Evaluate points
res.xs_iters.append(new_xs)
new_ys = [f(x) for x in new_xs]
res.ys_iters.append(new_ys)
total_evals += batch_size
known_xs.extend(new_xs)
known_ys.extend(new_ys)
# Save function value
if save_func_vals:
res.func_vals.append(f(current_x))
res.x_iters.append(np.copy(current_x))
res.fun = res.func_vals[-1]
# current sampling radius (maximal)
max_radius = 0
for x in new_xs:
if np.linalg.norm(x - current_x) > max_radius:
max_radius = np.linalg.norm(x - current_x)
reward = [-y for y in new_ys]
# warmup steps control whether to use the model to estimate the f
if num_iter >= warmup_steps:
# Determine points to use to build model
model_xs = []
model_ys = []
for x, y in zip(known_xs, known_ys):
if np.linalg.norm(x - current_x) < radius_coeff * max_radius:
model_xs.append(x)
model_ys.append(y)
# safer way without the `SVD` not converging
try:
model = _get_quadratic_model(model_xs, model_ys, x)
use_model = True
except ValueError:
use_model = False
if use_model:
# get samples (from model)
z = random_state.randn(batch_size_model, n)
new_xs = sigma * z + current_x
# use the model for prediction
new_ys = model.predict(new_xs - current_x)
reward = [-y for y in new_ys]
reward = np.array(reward)
# stats
reward_mean = np.mean(reward)
reward_max = np.max(reward)
if reward_max > history_max:
history_max = reward_max
# subtract baseline
reward = reward - reward_mean
# analytic derivatives (natural gradient policy gradient)
delta_mean = np.dot(z.T, reward) * sigma
delta_log_sigma = np.dot(z.T**2, reward) / np.sqrt(2)
delta_mean_norm = np.linalg.norm(np.dot(z.T, reward))
delta_log_sigma_norm = np.linalg.norm(np.dot(z.T**2, reward))
delta_mean = delta_mean / delta_mean_norm
delta_log_sigma = delta_log_sigma / delta_log_sigma_norm
# gradient ascend to update the parameters
current_x, m_mean, v_mean = _adam_update(delta_mean,
current_x,
num_iter,
m_mean,
v_mean,
lr_schedule=lr_schedule1)
log_sigma, m_log_sigma, v_log_sigma = _adam_update(
delta_log_sigma,
log_sigma,
num_iter,
m_log_sigma,
v_log_sigma,
lr_schedule=lr_schedule2,
)
log_sigma = np.clip(log_sigma, -20.0, 2.0)
sigma = np.exp(log_sigma)
num_iter += 1
final_val = f(current_x)
res.func_vals.append(final_val)
if message is None:
message = "Reached maximum number of iterations."
res.x_iters.append(current_x)
total_evals += 1
res.x = current_x
res.fun = final_val
res.nit = num_iter
res.nfev = total_evals
res.message = message
return res
def model_gradient_descent(
f: Callable[..., float],
x0: np.ndarray,
*,
args=(),
rate: float = 1e-1,
sample_radius: float = 1e-1,
n_sample_points: int = 100,
n_sample_points_ratio: Optional[float] = None,
rate_decay_exponent: float = 0.0,
stability_constant: float = 0.0,
sample_radius_decay_exponent: float = 0.0,
tol: float = 1e-8,
known_values: Optional[Tuple[List[np.ndarray], List[float]]] = None,
max_iterations: Optional[int] = None,
max_evaluations: Optional[int] = None) -> scipy.optimize.OptimizeResult:
"""Model gradient descent algorithm for black-box optimization.
The idea of this algorithm is to perform gradient descent, but estimate
the gradient using a surrogate model instead of, say, by
finite-differencing. The surrogate model is a least-squared quadratic
fit to points sampled from the vicinity of the current iterate.
This algorithm works well when you have an initial guess which is in the
convex neighborhood of a local optimum and you want to converge to that
local optimum. It's meant to be used when the function is stochastic.
Args:
f: The function to minimize.
x0: An initial guess.
args: Additional arguments to pass to the function.
rate: The learning rate for the gradient descent.
sample_radius: The radius around the current iterate to sample
points from to build the quadratic model.
n_sample_points: The number of points to sample in each iteration.
n_sample_points_ratio: This specifies the number of points to sample
in each iteration as a coefficient of the number of points
required to exactly determine a quadratic model. The number
of sample points will be this coefficient times (n+1)(n+2)/2,
rounded up, where n is the number of parameters.
Setting this overrides n_sample_points.
rate_decay_exponent: Controls decay of learning rate.
In each iteration, the learning rate is changed to the
base learning rate divided by (i + 1 + S)**a, where S
is the stability constant and a is the rate decay exponent
(this parameter).
stability_constant: Affects decay of learning rate.
In each iteration, the learning rate is changed to the
base learning rate divided by (i + 1 + S)**a, where S
is the stability constant (this parameter) and a is the rate decay
exponent.
sample_radius_decay_exponent: Controls decay of sample radius.
tol: The algorithm terminates when the difference between the current
iterate and the next suggested iterate is smaller than this value.
known_values: Any prior known values of the objective function.
This is given as a tuple where the first element is a list
of points and the second element is a list of the function values
at those points.
max_iterations: The maximum number of iterations to allow before
termination.
max_evaluations: The maximum number of function evaluations to allow
before termination.
Returns:
Scipy OptimizeResult
"""
if known_values is not None:
known_xs, known_ys = known_values
known_xs = [np.copy(x) for x in known_xs]
known_ys = [np.copy(y) for y in known_ys]
else:
known_xs, known_ys = [], []
if max_iterations is None:
max_iterations = np.inf
if max_evaluations is None:
max_evaluations = np.inf
n = len(x0)
if n_sample_points_ratio is not None:
n_sample_points = int(
np.ceil(n_sample_points_ratio * (n + 1) * (n + 2) / 2))
_, f = wrap_function(f, args)
res = OptimizeResult()
current_x = np.copy(x0)
res.x_iters = [] # initializes as lists
res.xs_iters = []
res.ys_iters = []
res.func_vals = []
res.model_vals = [None]
res.fun = 0
total_evals = 0
num_iter = 0
converged = False
message = None
while num_iter < max_iterations:
current_sample_radius = (sample_radius /
(num_iter + 1)**sample_radius_decay_exponent)
# Determine points to evaluate
# in ball around current point
new_xs = [np.copy(current_x)] + [
current_x + _random_point_in_ball(n, current_sample_radius)
for _ in range(n_sample_points)
]
if total_evals + len(new_xs) > max_evaluations:
message = 'Reached maximum number of evaluations.'
break
# Evaluate points
res.xs_iters.append(new_xs)
new_ys = [f(x) for x in new_xs]
res.ys_iters.append(new_ys)
total_evals += len(new_ys)
known_xs.extend(new_xs)
known_ys.extend(new_ys)
# Save function value
res.func_vals.append(new_ys[0])
res.x_iters.append(np.copy(current_x))
res.fun = res.func_vals[-1]
# Determine points to use to build model
model_xs = []
model_ys = []
for x, y in zip(known_xs, known_ys):
if np.linalg.norm(x - current_x) < current_sample_radius:
model_xs.append(x)
model_ys.append(y)
# Build and solve model
model_gradient, model = _get_least_squares_model_gradient(
model_xs, model_ys, current_x)
# calculate the gradient and update the current point
gradient_norm = np.linalg.norm(model_gradient)
decayed_rate = (
rate / (num_iter + 1 + stability_constant)**rate_decay_exponent)
# Convergence criteria
if decayed_rate * gradient_norm < tol:
converged = True
message = 'Optimization converged successfully.'
break
# Update
current_x -= decayed_rate * model_gradient
res.model_vals.append(
model.predict([-decayed_rate * model_gradient])[0])
num_iter += 1
if converged:
final_val = res.func_vals[-1]
else:
final_val = f(current_x)
res.func_vals.append(final_val)
if message is None:
message = 'Reached maximum number of iterations.'
res.x_iters.append(current_x)
total_evals += 1
res.x = current_x
res.fun = final_val
res.nit = num_iter
res.nfev = total_evals
res.message = message
return res
def dual_annealing(func, x0, bounds, args=(), maxiter=1000,
local_search_options={}, initial_temp=5230.,
restart_temp_ratio=2.e-5, visit=2.62, accept=-5.0,
maxfun=1e7, seed=None, no_local_search=False,
callback=None):
"""
Find the global minimum of a function using Dual Annealing.
Parameters
----------
func : callable
The objective function to be minimized. Must be in the form
``f(x, *args)``, where ``x`` is the argument in the form of a 1-D array
and ``args`` is a tuple of any additional fixed parameters needed to
completely specify the function.
x0 : ndarray, shape(n,)
A single initial starting point coordinates. If ``None`` is provided,
initial coordinates are automatically generated (using the ``reset``
method from the internal ``EnergyState`` class).
bounds : sequence, shape (n, 2)
Bounds for variables. ``(min, max)`` pairs for each element in ``x``,
defining bounds for the objective function parameter.
args : tuple, optional
Any additional fixed parameters needed to completely specify the
objective function.
maxiter : int, optional
The maximum number of global search iterations. Default value is 1000.
local_search_options : dict, optional
Extra keyword arguments to be passed to the local minimizer
(`minimize`). Some important options could be:
``method`` for the minimizer method to use and ``args`` for
objective function additional arguments.
initial_temp : float, optional
The initial temperature, use higher values to facilitates a wider
search of the energy landscape, allowing dual_annealing to escape
local minima that it is trapped in. Default value is 5230. Range is
(0.01, 5.e4].
restart_temp_ratio : float, optional
During the annealing process, temperature is decreasing, when it
reaches ``initial_temp * restart_temp_ratio``, the reannealing process
is triggered. Default value of the ratio is 2e-5. Range is (0, 1).
visit : float, optional
Parameter for visiting distribution. Default value is 2.62. Higher
values give the visiting distribution a heavier tail, this makes
the algorithm jump to a more distant region. The value range is (0, 3].
accept : float, optional
Parameter for acceptance distribution. It is used to control the
probability of acceptance. The lower the acceptance parameter, the
smaller the probability of acceptance. Default value is -5.0 with
a range (-1e4, -5].
maxfun : int, optional
Soft limit for the number of objective function calls. If the
algorithm is in the middle of a local search, this number will be
exceeded, the algorithm will stop just after the local search is
done. Default value is 1e7.
seed : {int or `numpy.random.RandomState` instance}, optional
If `seed` is not specified the `numpy.random.RandomState` singleton is
used.
If `seed` is an int, a new ``RandomState`` instance is used,
seeded with `seed`.
If `seed` is already a ``RandomState`` instance, then that
instance is used.
Specify `seed` for repeatable minimizations. The random numbers
generated with this seed only affect the visiting distribution
function and new coordinates generation.
no_local_search : bool, optional
If `no_local_search` is set to True, a traditional Generalized
Simulated Annealing will be performed with no local search
strategy applied.
callback : callable, optional
A callback function with signature ``callback(x, f, context)``,
which will be called for all minima found.
``x`` and ``f`` are the coordinates and function value of the
latest minimum found, and ``context`` has value in [0, 1, 2], with the
following meaning:
- 0: minimum detected in the annealing process.
- 1: detection occured in the local search process.
- 2: detection done in the dual annealing process.
If the callback implementation returns True, the algorithm will stop.
Returns
-------
res : OptimizeResult
The optimization result represented as a `OptimizeResult` object.
Important attributes are: ``x`` the solution array, ``fun`` the value
of the function at the solution, and ``message`` which describes the
cause of the termination.
See `OptimizeResult` for a description of other attributes.
Notes
-----
This function implements the Dual Annealing optimization. This stochastic
approach derived from [3]_ combines the generalization of CSA (Classical
Simulated Annealing) and FSA (Fast Simulated Annealing) [1]_ [2]_ coupled
to a strategy for applying a local search on accepted locations [4]_.
An alternative implementation of this same algorithm is described in [5]_
and benchmarks are presented in [6]_. This approach introduces an advanced
method to refine the solution found by the generalized annealing
process. This algorithm uses a distorted Cauchy-Lorentz visiting
distribution, with its shape controlled by the parameter :math:`q_{v}`
.. math::
g_{q_{v}}(\\Delta x(t)) \\propto \\frac{ \\
\\left[T_{q_{v}}(t) \\right]^{-\\frac{D}{3-q_{v}}}}{ \\
\\left[{1+(q_{v}-1)\\frac{(\\Delta x(t))^{2}} { \\
\\left[T_{q_{v}}(t)\\right]^{\\frac{2}{3-q_{v}}}}}\\right]^{ \\
\\frac{1}{q_{v}-1}+\\frac{D-1}{2}}}
Where :math:`t` is the artificial time. This visiting distribution is used
to generate a trial jump distance :math:`\\Delta x(t)` of variable
:math:`x(t)` under artificial temperature :math:`T_{q_{v}}(t)`.
From the starting point, after calling the visiting distribution
function, the acceptance probability is computed as follows:
.. math::
p_{q_{a}} = \\min{\\{1,\\left[1-(1-q_{a}) \\beta \\Delta E \\right]^{ \\
\\frac{1}{1-q_{a}}}\\}}
Where :math:`q_{a}` is a acceptance parameter. For :math:`q_{a}<1`, zero
acceptance probability is assigned to the cases where
.. math::
[1-(1-q_{a}) \\beta \\Delta E] < 0
The artificial temperature :math:`T_{q_{v}}(t)` is decreased according to
.. math::
T_{q_{v}}(t) = T_{q_{v}}(1) \\frac{2^{q_{v}-1}-1}{\\left( \\
1 + t\\right)^{q_{v}-1}-1}
Where :math:`q_{v}` is the visiting parameter.
.. versionadded:: 1.2.0
References
----------
.. [1] Tsallis C. Possible generalization of Boltzmann-Gibbs
statistics. Journal of Statistical Physics, 52, 479-487 (1998).
.. [2] Tsallis C, Stariolo DA. Generalized Simulated Annealing.
Physica A, 233, 395-406 (1996).
.. [3] Xiang Y, Sun DY, Fan W, Gong XG. Generalized Simulated
Annealing Algorithm and Its Application to the Thomson Model.
Physics Letters A, 233, 216-220 (1997).
.. [4] Xiang Y, Gong XG. Efficiency of Generalized Simulated
Annealing. Physical Review E, 62, 4473 (2000).
.. [5] Xiang Y, Gubian S, Suomela B, Hoeng J. Generalized
Simulated Annealing for Efficient Global Optimization: the GenSA
Package for R. The R Journal, Volume 5/1 (2013).
.. [6] Mullen, K. Continuous Global Optimization in R. Journal of
Statistical Software, 60(6), 1 - 45, (2014). DOI:10.18637/jss.v060.i06
Examples
--------
The following example is a 10-dimensional problem, with many local minima.
The function involved is called Rastrigin
(https://en.wikipedia.org/wiki/Rastrigin_function)
>>> from scipy.optimize import dual_annealing
>>> func = lambda x: np.sum(x*x - 10*np.cos(2*np.pi*x)) + 10*np.size(x)
>>> lw = [-5.12] * 10
>>> up = [5.12] * 10
>>> ret = dual_annealing(func, None, bounds=list(zip(lw, up)), seed=1234)
>>> print("global minimum: xmin = {0}, f(xmin) = {1:.6f}".format(
... ret.x, ret.fun))
global minimum: xmin = [-4.26437714e-09 -3.91699361e-09 -1.86149218e-09 -3.97165720e-09
-6.29151648e-09 -6.53145322e-09 -3.93616815e-09 -6.55623025e-09
-6.05775280e-09 -5.00668935e-09], f(xmin) = 0.000000
"""
if x0 is not None and not len(x0) == len(bounds):
raise ValueError('Bounds size does not match x0')
lu = list(zip(*bounds))
lower = np.array(lu[0])
upper = np.array(lu[1])
# Check that restart temperature ratio is correct
if restart_temp_ratio <= 0. or restart_temp_ratio >= 1.:
raise ValueError('Restart temperature ratio has to be in range (0, 1)')
# Checking bounds are valid
if (np.any(np.isinf(lower)) or np.any(np.isinf(upper)) or np.any(
np.isnan(lower)) or np.any(np.isnan(upper))):
raise ValueError('Some bounds values are inf values or nan values')
# Checking that bounds are consistent
if not np.all(lower < upper):
raise ValueError('Bounds are note consistent min < max')
# Wrapper for the objective function
func_wrapper = ObjectiveFunWrapper(func, maxfun, *args)
# Wrapper fot the minimizer
minimizer_wrapper = LocalSearchWrapper(
bounds, func_wrapper, **local_search_options)
# Initialization of RandomState for reproducible runs if seed provided
rand_state = check_random_state(seed)
# Initialization of the energy state
energy_state = EnergyState(lower, upper, callback)
energy_state.reset(func_wrapper, rand_state, x0)
# Minimum value of annealing temperature reached to perform
# re-annealing
temperature_restart = initial_temp * restart_temp_ratio
# VisitingDistribution instance
visit_dist = VisitingDistribution(lower, upper, visit, rand_state)
# Strategy chain instance
strategy_chain = StrategyChain(accept, visit_dist, func_wrapper,
minimizer_wrapper, rand_state, energy_state)
# Run the search loop
need_to_stop = False
iteration = 0
message = []
t1 = np.exp((visit - 1) * np.log(2.0)) - 1.0
while(not need_to_stop):
for i in range(maxiter):
# Compute temperature for this step
s = float(i) + 2.0
t2 = np.exp((visit - 1) * np.log(s)) - 1.0
temperature = initial_temp * t1 / t2
iteration += 1
if iteration >= maxiter:
message.append("Maximum number of iteration reached")
need_to_stop = True
break
# Need a re-annealing process?
if temperature < temperature_restart:
energy_state.reset(func_wrapper, rand_state)
break
# starting strategy chain
val = strategy_chain.run(i, temperature)
if val is not None:
message.append(val)
need_to_stop = True
break
# Possible local search at the end of the strategy chain
if not no_local_search:
val = strategy_chain.local_search()
if val is not None:
message.append(val)
need_to_stop = True
break
# Return the OptimizeResult
res = OptimizeResult()
res.x = energy_state.xbest
res.fun = energy_state.ebest
res.nit = iteration
res.nfev = func_wrapper.nfev
res.njev = func_wrapper.ngev
res.message = message
return res
def optimize_minimize_mhmcmc_cluster(objective,
bounds,
args=(),
x0=None,
T=1,
N=3,
burnin=100000,
maxiter=1000000,
target_ar=0.4,
ar_tolerance=0.05,
cluster_eps=DEFAULT_CLUSTER_EPS,
rnd_seed=None,
collect_samples=None,
logger=None):
"""
Minimize objective function and return up to N local minima solutions.
:param objective: Objective function to minimize. Takes unpacked args as function call arguments and returns
a float.
:type objective: Callable(\*args) -> float
:param bounds: Bounds of the parameter space.
:type bounds: scipy.optimize.Bounds
:param args: Any additional fixed parameters needed to completely specify the objective function.
:type args: tuple or list
:param x0: Initial guess. If None, will be selected randomly and uniformly within the parameter bounds.
:type x0: numpy.array with same shape as elements of bounds
:param T: The "temperature" parameter for the accept or reject criterion. To sample the domain well,
should be in the order of the typical difference in local minima objective valuations.
:type T: float
:param N: Maximum number of minima to return
:type N: int
:param burnin: Number of random steps to discard before starting to accumulate statistics.
:type burnin: int
:param maxiter: Maximum number of steps to take (including burnin).
:type maxiter: int
:param target_ar: Target acceptance rate of point samples generated by stepping.
:type target_ar: float between 0 and 1
:param ar_tolerance: Tolerance on the acceptance rate before actively adapting the step size.
:type ar_tolerance: float
:param cluster_eps: Point proximity tolerance for DBSCAN clustering, in normalized bounds coordinates.
:type cluster_eps: float
:param rnd_seed: Random seed to force deterministic behaviour
:type rnd_seed: int
:param collect_samples: If not None and integral type, collect collect_samples at regular intervals
and return as part of solution.
:type collect_samples: int or NoneType
:param logger: Logger instance for outputting log messages.
:return: OptimizeResult containing solution(s) and solver data.
:rtype: scipy.optimize.OptimizeResult with additional attributes
"""
@call_counter
def obj_counted(*args):
return objective(*args)
# end func
assert maxiter >= 2 * burnin, "maxiter {} should be at least twice burnin steps {}".format(
maxiter, burnin)
main_iter = maxiter - burnin
if collect_samples is not None:
assert isinstance(collect_samples,
int), "collect_samples expected to be integral type"
assert collect_samples > 0, "collect_samples expected to be positive"
# end if
beta = 1.0 / T
if rnd_seed is None:
rnd_seed = int(time.time() * 1000) % (1 << 31)
# end if
np.random.seed(rnd_seed)
if logger:
logger.info('Using random seed {}'.format(rnd_seed))
# end
if x0 is None:
x0 = np.random.uniform(bounds.lb, bounds.ub)
# end if
assert np.all((x0 >= bounds.lb) & (x0 <= bounds.ub))
x = x0.copy()
funval = obj_counted(x, *args)
# Set up stepper with adaptive acceptance rate
stepper = BoundedRandNStepper(bounds)
stepper = AdaptiveStepsize(stepper,
accept_rate=target_ar,
ar_tolerance=ar_tolerance,
interval=50)
# -------------------------------
# DO BURN-IN
rejected_randomly = 0
accepted_burnin = 0
tracked_range = tqdm(range(burnin), total=burnin, desc='BURN-IN')
if logger:
stepper.logger = lambda msg: tracked_range.write(logger.name + ':' +
msg)
else:
stepper.logger = tracked_range.write
# end if
for _ in tracked_range:
x_new = stepper(x)
funval_new = obj_counted(x_new, *args)
log_alpha = -(funval_new - funval) * beta
if log_alpha > 0 or np.log(np.random.rand()) <= log_alpha:
x = x_new
funval = funval_new
stepper.notify_accept()
accepted_burnin += 1
elif log_alpha <= 0:
rejected_randomly += 1
# end if
# end for
ar = float(accepted_burnin) / burnin
if logger:
logger.info("Burn-in acceptance rate: {}".format(ar))
# end if
# -------------------------------
# DO MAIN LOOP
if collect_samples is not None:
nsamples = min(collect_samples, main_iter)
sample_cadence = main_iter / nsamples
samples = np.zeros((nsamples, len(x)))
samples_fval = np.zeros(nsamples)
# end if
accepted = 0
rejected_randomly = 0
minima_sorted = SortedList(
key=lambda rec: rec[1]) # Sort by objective function value
hist = HistogramIncremental(bounds, nbins=100)
# Cached a lot of potential minimum values, as these need to be clustered before return N results
N_cached = int(np.ceil(N * main_iter / 500))
next_sample = 0.0
sample_count = 0
tracked_range = tqdm(range(main_iter), total=main_iter, desc='MAIN')
if logger:
stepper.logger = lambda msg: tracked_range.write(logger.name + ':' +
msg)
else:
stepper.logger = tracked_range.write
# end if
for i in tracked_range:
if collect_samples and i >= next_sample:
assert sample_count < collect_samples
samples[sample_count] = x
samples_fval[sample_count] = funval
sample_count += 1
next_sample += sample_cadence
# end if
x_new = stepper(x)
funval_new = obj_counted(x_new, *args)
log_alpha = -(funval_new - funval) * beta
if log_alpha > 0 or np.log(np.random.rand()) <= log_alpha:
x = x_new
funval = funval_new
minima_sorted.add((x, funval))
if len(minima_sorted) > N_cached:
minima_sorted.pop()
# end if
stepper.notify_accept()
hist += x
accepted += 1
elif log_alpha <= 0:
rejected_randomly += 1
# end if
# end for
stepper.logger = None
ar = float(accepted) / main_iter
if logger:
logger.info("Acceptance rate: {}".format(ar))
logger.info("Best minima (before clustering):\n{}".format(
np.array([_mx[0] for _mx in minima_sorted[:10]])))
# end if
# -------------------------------
# Cluster minima and associate each cluster with a local minimum.
# Using a normalized coordinate space for cluster detection.
x_range = bounds.ub - bounds.lb
pts = np.array([x[0] for x in minima_sorted])
fvals = np.array([x[1] for x in minima_sorted])
pts_norm = (pts - bounds.lb) / x_range
_, labels = dbscan(pts_norm, eps=cluster_eps, min_samples=21, n_jobs=-1)
# Compute mean of each cluster and evaluate objective function at cluster mean locations.
minima_candidates = []
for grp in range(max(labels) + 1):
mask = (labels == grp)
mean_loc = np.mean(pts[mask, :], axis=0)
# Evaluate objective function precisely at the mean location of each cluster
fval = obj_counted(mean_loc, *args)
minima_candidates.append((mean_loc, grp, fval))
# end for
# Rank minima locations by objective function.
minima_candidates.sort(key=lambda c: c[2])
# Pick up to N solutions
solutions = minima_candidates[:N]
# Put results into OptimizeResult container.
# Add histograms to output result (in form of scipy.stats.rv_histogram)
solution = OptimizeResult()
solution.x = np.array([s[0] for s in solutions])
solution.clusters = [pts[(labels == s[1])] for s in solutions]
solution.cluster_funvals = [fvals[(labels == s[1])] for s in solutions]
solution.bins = hist.bins
solution.distribution = hist.histograms
solution.acceptance_rate = ar
solution.success = True
solution.status = 0
if len(solutions) > 0:
solution.message = 'SUCCESS: Found {} local minima'.format(
len(solutions))
else:
solution.message = 'WARNING: Found no clusters within tolerance {}'.format(
cluster_eps)
# end if
solution.fun = np.array([s[2] for s in solutions])
solution.jac = None
solution.nfev = obj_counted.counter
solution.njev = 0
solution.nit = main_iter
solution.maxcv = None
solution.samples = samples if collect_samples else None
solution.sample_funvals = samples_fval if collect_samples else None
solution.bounds = bounds
solution.version = 's0.3' # Solution version for future traceability
solution.rnd_seed = rnd_seed
return solution
