def advance(self, evaluator: Evaluator):
self.iteration += 1
logger.info("Starting SPSA iteration %d." % self.iteration)
if self.parameters is None:
self.parameters = evaluator.problem.initial
# Calculate objective
if self.compute_objective:
annotations = { "type": "objective" }
objective_identifier = evaluator.submit(self.parameters, annotations = annotations)
# Update step lengths
gradient_length = self.gradient_factor / (self.iteration + self.gradient_offset)**self.gradient_exponent
perturbation_length = self.perturbation_factor / self.iteration**self.perturbation_exponent
# Sample direction from Rademacher distribution
direction = self.random.randint(0, 2, len(self.parameters)) - 0.5
annotations = {
"gradient_length": gradient_length,
"perturbation_length": perturbation_length,
"direction": direction,
"type": "gradient"
}
# Schedule samples
positive_parameters = np.copy(self.parameters)
positive_parameters += direction * perturbation_length
annotations = deep_merge.merge(annotations, { "type": "positive_gradient" })
positive_identifier = evaluator.submit(positive_parameters, annotations = annotations)
negative_parameters = np.copy(self.parameters)
negative_parameters -= direction * perturbation_length
annotations = deep_merge.merge(annotations, { "type": "negative_gradient" })
negative_identifier = evaluator.submit(negative_parameters, annotations = annotations)
# Wait for gradient run results
evaluator.wait()
if self.compute_objective:
evaluator.clean(objective_identifier)
positive_objective, positive_state = evaluator.get(positive_identifier)
evaluator.clean(positive_identifier)
negative_objective, negative_state = evaluator.get(negative_identifier)
evaluator.clean(negative_identifier)
g_k = (positive_objective - negative_objective) / (2.0 * perturbation_length)
g_k *= direction**-1
# Update state
self.parameters -= gradient_length * g_k
def advance(self, evaluator: Evaluator):
self.iteration += 1
logger.info("Starting Random Walk iteration %d" % self.iteration)
parameters = [np.array([
bounds[0] + self.random.random() * (bounds[1] - bounds[0]) # TODO: Not demterinistic!
for bounds in self.bounds
]) for k in range(self.parallel)]
identifiers = [evaluator.submit(p) for p in parameters]
evaluator.wait(identifiers)
evaluator.clean()
def advance(self, evaluator: Evaluator):
if self.iteration == 0:
logger.info("Initializing Opdyts")
self.initial_identifier = evaluator.submit(self.initial_parameters,
{ "iterations": 1 }, { "type": "initial", "transient": True }
)
self.initial_objective, self.initial_state = evaluator.get(self.initial_identifier)
self.iteration += 1
logger.info("Starting Opdyts iteration %d" % self.iteration)
# Create new set of candidate parameters
candidate_parameters = np.zeros((self.candidate_set_size, self.number_of_parameters))
for c in range(0, self.candidate_set_size, 2):
direction = self.random.random_sample(size = (self.number_of_parameters,)) * 2.0 - 1.0
candidate_parameters[c] = self.initial_parameters + direction * self.perturbation_length
candidate_parameters[c + 1] = self.initial_parameters + direction * self.perturbation_length
# Find initial candiate states
candidate_identifiers = []
candidate_states = np.zeros((self.candidate_set_size, self.number_of_states))
candidate_deltas = np.zeros((self.candidate_set_size, self.number_of_states))
candidate_objectives = np.zeros((self.candidate_set_size,))
candidate_transitions = np.ones((self.candidate_set_size,))
annotations = {
"type": "candidate", "v": self.v, "w": self.w,
"transient": True, "iteration": self.iteration
}
for c in range(self.candidate_set_size):
candidate_annotations = copy.copy(annotations)
candidate_annotations.update({ "candidate": c })
candidate_identifiers.append(evaluator.submit(candidate_parameters[c], {
#"iterations": transition_iterations,
"restart": self.initial_identifier
}, candidate_annotations))
evaluator.wait()
for c in range(self.candidate_set_size):
candidate_objectives[c], candidate_states[c] = evaluator.get(candidate_identifiers[c])
candidate_deltas[c] = candidate_states[c] - self.initial_state
# Advance candidates
local_adaptation_transient_performance = []
local_adaptation_equilibrium_gap = []
local_adaptation_uniformity_gap = []
while np.max(candidate_transitions) < self.number_of_transitions:
# Approximate selection problem
selection_problem = ApproximateSelectionProblem(self.v, self.w, candidate_deltas, candidate_objectives)
alpha = selection_problem.solve()
transient_performance = selection_problem.get_transient_performance(alpha)
equilibrium_gap = selection_problem.get_equilibrium_gap(alpha)
uniformity_gap = selection_problem.get_uniformity_gap(alpha)
local_adaptation_transient_performance.append(transient_performance)
local_adaptation_equilibrium_gap.append(equilibrium_gap)
local_adaptation_uniformity_gap.append(uniformity_gap)
logger.info(
"Transient performance: %f, Equilibirum gap: %f, Uniformity_gap: %f",
transient_performance, equilibrium_gap, uniformity_gap)
cumulative_alpha = np.cumsum(alpha)
c = np.sum(self.random.random_sample() > cumulative_alpha) # TODO: Not deterministic!
logger.info("Transitioning candidate %d", c)
candidate_transitions[c] += 1
transient = candidate_transitions[c] < self.number_of_transitions
annotations.update({
"type": "transition",
"candidate": c, "transient_performance": transient_performance,
"equilibrium_gap": equilibrium_gap, "uniformity_gap": uniformity_gap,
"transient": transient
})
# Advance selected candidate
identifier = evaluator.submit(candidate_parameters[c], {
#"iterations": transition_iterations,
"restart": candidate_identifiers[c]
}, annotations)
new_objective, new_state = evaluator.get(identifier)
evaluator.clean(candidate_identifiers[c])
candidate_deltas[c] = new_state - candidate_states[c]
candidate_states[c], candidate_objectives[c] = new_state, new_objective
candidate_identifiers[c] = identifier
index = np.argmax(candidate_transitions)
logger.info("Solved selection problem with candidate %d", index)
for c in range(self.candidate_set_size):
if c != index:
evaluator.clean(candidate_identifiers[c])
evaluator.clean(self.initial_identifier)
self.initial_identifier = candidate_identifiers[index]
self.initial_state = candidate_states[index]
self.initial_parameters = candidate_parameters[index]
self.adaptation_selection_performance.append(candidate_objectives[index])
self.adaptation_transient_performance.append(np.array(local_adaptation_transient_performance))
self.adaptation_equilibrium_gap.append(np.array(local_adaptation_equilibrium_gap))
self.adaptation_uniformity_gap.append(np.array(local_adaptation_uniformity_gap))
adaptation_problem = AdaptationProblem(self.adaptation_weight, self.adaptation_selection_performance, self.adaptation_transient_performance, self.adaptation_equilibrium_gap, self.adaptation_uniformity_gap)
self.v, self.w = adaptation_problem.solve()
logger.info("Solved Adaptation Problem. v = %f, w = %f", self.v, self.w)
def advance(self, evaluator: Evaluator):
if self.iteration == 0:
self.mean = np.copy(self.initial_values).reshape((self.N, 1))
self.iteration += 1
logger.info("Starting CMA-ES iteration %d." % self.iteration)
annotations = {
"mean": self.mean,
"covariance": self.C,
"pc": self.pc,
"ps": self.ps,
"sigma": self.sigma
}
self.counteval += self.L
candidate_parameters = self.sigma * np.dot(
(self.random.normal(size=(self.N, self.L)) *
self.D[:, np.newaxis]).T, self.B) + self.mean.T
candidate_identifiers = [
evaluator.submit(parameters, annotations=annotations)
for parameters in candidate_parameters
]
# Wait for samples
evaluator.wait()
# Obtain fitness
candidate_objectives = np.array([
evaluator.get(identifier)[0] # We minimize!
for identifier in candidate_identifiers
])
# Cleanup
for identifier in candidate_identifiers:
evaluator.clean(identifier)
sorter = np.argsort(candidate_objectives)
candidate_objectives = candidate_objectives[sorter]
candidate_parameters = candidate_parameters[sorter, :]
# Update mean
previous_mean = self.mean
self.mean = np.sum(candidate_parameters[:self.mu] *
self.weights[:, np.newaxis],
axis=0).reshape((self.N, 1))
# Update evolution paths
psa = (1.0 - self.cs) * self.ps
psb = np.sqrt(self.cs * (2.0 - self.cs) * self.mueff) * np.dot(
self.invsqrtC, self.mean - previous_mean) / self.sigma
self.ps = psa + psb
hsig = la.norm(
self.ps) / np.sqrt(1.0 -
(1.0 - self.cs)**(2.0 * self.counteval / self.L)
) / self.chiN < 1.4 + 2.0 / (self.N + 1.0)
pca = (1.0 - self.cc) * self.pc
pcb = hsig * np.sqrt(self.cc * (2.0 - self.cc) * self.mueff) * (
self.mean - previous_mean) / self.sigma
self.pc = pca + pcb
# Adapt covariance matrix
artmp = (1.0 / self.sigma) * (candidate_parameters[:self.mu].T -
previous_mean)
Ca = (1.0 - self.c1 - self.cmu) * self.C
Cb = self.c1 * (np.dot(self.pc, self.pc.T) + (not hsig) * self.cc *
(2.0 - self.cc) * self.C)
Cc = self.cmu * np.dot(artmp, np.dot(np.diag(self.weights), artmp.T))
C = Ca + Cb + Cc
# Adapt step size
self.sigma = self.sigma * np.exp(
(self.cs / self.damps) * (la.norm(self.ps) / self.chiN - 1.0))
if self.counteval - self.eigeneval > self.L / (
self.c1 + self.cmu) / self.N / 10.0:
self.eigeneval = self.counteval
self.C = np.triu(self.C) + np.triu(self.C, 1).T
d, self.B = la.eig(self.C)
self.D = np.sqrt(d)
Dm = np.diag(1.0 / np.sqrt(d))
self.invsqrtC = np.dot(self.B.T, np.dot(Dm, self.B))
if np.max(self.D) > 1e7 * np.min(self.D):
logger.warning("Condition exceeds 1e14")
def advance(self, evaluator: Evaluator):
self.iteration += 1
logger.info("Starting FDSA iteration %d." % self.iteration)
# Calculate objective
if self.compute_objective:
annotations = {"type": "objective"}
objective_identifier = evaluator.submit(self.parameters,
annotations=annotations)
# Update lengths
gradient_length = self.gradient_factor / (
self.iteration + self.gradient_offset)**self.gradient_exponent
perturbation_length = self.perturbation_factor / self.iteration**self.perturbation_exponent
annotations = {
"gradient_length": gradient_length,
"perturbation_length": perturbation_length,
"type": "gradient"
}
# I) Calculate gradients
gradient = np.zeros((len(self.parameters), ))
gradient_information = []
# Schedule all necessary runs
for d in range(len(self.parameters)):
annotations = deep_merge.merge(annotations, {"dimension": d})
positive_parameters = np.copy(self.parameters)
positive_parameters[d] += perturbation_length
annotations = deep_merge.merge(annotations,
{"type": "positive_gradient"})
positive_identifier = evaluator.submit(positive_parameters,
annotations=annotations)
negative_parameters = np.copy(self.parameters)
negative_parameters[d] -= perturbation_length
annotations = deep_merge.merge(annotations,
{"sign": "negative_gradient"})
negative_identifier = evaluator.submit(negative_parameters,
annotations=annotations)
gradient_information.append(
(positive_parameters, positive_identifier, negative_parameters,
negative_identifier))
# Wait for gradient run results
evaluator.wait()
if self.compute_objective:
evaluator.clean(objective_identifier)
for d, item in enumerate(gradient_information):
positive_parameters, positive_identifier, negative_parameters, negative_identifier = item
positive_objective, positive_state = evaluator.get(
positive_identifier)
evaluator.clean(positive_identifier)
negative_objective, negative_state = evaluator.get(
negative_identifier)
evaluator.clean(negative_identifier)
gradient[d] = (positive_objective -
negative_objective) / (2.0 * perturbation_length)
# II) Update state
self.parameters -= gradient_length * gradient