def fit(self, data):
# Resets the rule base
self.rule_base = rbm.init_rule_base(self.fuzzy_sets, self.order)
# Compute the universe of discourse
data_range = np.max(data) - np.min(data)
if self.bound_type == 'min-max':
lb = np.min(data) - data_range * 0.1
ub = np.max(data) + data_range * 0.1
else: # self.bound_type == 'mu-sigma'
lb = np.mean(data) - np.std(data) * self.sigma_multiplier
ub = np.mean(data) + np.std(data) * self.sigma_multiplier
# Regenerate partitions
if self.partitionner == 'uniform':
self.partitions = pu.generate_t_partitions(self.nsets, lb, ub)
else: # self.bound_type == 'knn'
self.partitions = pu.generate_t_partitions_knn(data, self.nsets, lb, ub)
# Populate the rule base
for i in range(len(data)-self.order):
window = data[i:(i+self.order+1)]
# Get pertinences
pertinence_list = pf.t_pertinence_list(window, self.partitions)
# Build rule
rule = rbm.generate_rule(pertinence_list)
# Add rule to the rule base
rbm.add_rule(self.rule_base, rule, self.nsets, self.order)
def retrain(self, data):
# Resets the rule base
self.rule_base = rbm.init_rule_base(self.fuzzy_sets, self.order)
# Regenerate partitions
data_range = np.max(data) - np.min(data)
if self.bound_type == 'min-max':
self.partitions = pu.generate_t_partitions(
self.nsets,
np.min(data) - data_range * 0.1,
np.max(data) + data_range * 0.1)
else: # self.bound_type == 'mu-sigma'
self.partitions = pu.generate_t_partitions(
self.nsets,
np.mean(data) - np.std(data) * self.sigma_multiplier,
np.mean(data) + np.std(data) * self.sigma_multiplier)
# Populate the rule base
for i in range(len(data) - self.order):
window = data[i:(i + self.order + 1)]
# Get pertinences
pertinence_list = pf.t_pertinence_list(window, self.partitions)
# Build rule
rule = rbm.generate_rule(pertinence_list)
# Add rule to the rule base
rbm.add_rule(self.rule_base, rule, self.nsets, self.order)
def fit(self, data):
x = data[-1]
# Verify the Universe of discourse
if self.bound_type == 'min-max':
data_range = self.max_val - self.min_val
lb = self.min_val - data_range * 0.1
ub = self.max_val + data_range * 0.1
elif self.bound_type == 'mu-sigma':
lb = self.mu - self.sigma * self.sigma_multiplier
ub = self.mu + self.sigma * self.sigma_multiplier
if not self.partitions:
self.partitions = pu.generate_t_partitions(self.nsets, lb, ub)
self.centers = [c[1] for c in self.partitions]
else: # Update centers
if self.partitionner == 'knn':
# Find the closest center to data
closest = np.argmin([np.abs(x - c)] for c in self.centers)
# Update the center
self.centers[closest] = ((self.centers[closest] * self.samples_per_cluster[closest]) + x) / \
(self.samples_per_cluster[closest] + 1)
self.samples_per_cluster[closest] += 1
elif self.partitionner == 'uniform':
mock_partitions = pu.generate_t_partitions(self.nsets, lb, ub)
self.centers = [c[1] for c in mock_partitions]
# Now we have two options
# 1) Generate new partitions and translate rules
if self.mod == 'translate':
self.translate()
else: # 2) Generate new partitions and do nothing
self.partitions = pu.generate_t_partitions_from_centers(
self.centers)
# Now we have three options
# 1) Do nothing
# 2) Delete old rules
# 3) Delete bad rules
if self.del_strategy == 'delete_bad':
self.delete_bad(x)
# elif self.del_strategy == 'delete_old':
# pass
# self.delete_old()
# Get pertinences
pertinence_list = pf.t_pertinence_list(data, self.partitions)
# Build rule
rule = rbm.generate_rule(pertinence_list)
# Add rule to the rule base
rbm.add_rule(self.rule_base, rule, self.nsets, self.order)
def fit(self, data):
for i in range(len(data) - self.order):
window = data[i:(i + self.order + 1)]
# Get pertinences
pertinence_list = pf.t_pertinence_list(window, self.partitions)
# Build rule
rule = rbm.generate_rule(pertinence_list)
# Add rule to the rule base
rbm.add_rule(self.rule_base, rule, self.nsets, self.order)
def translate(self, data):
if not self.rule_base:
self.rule_base = rbm.init_rule_base(self.fuzzy_sets, self.order)
if self.bound_type == 'min-max':
data_range = np.max(data) - np.min(data)
lb = np.min(data) - data_range * 0.5
ub = np.max(data) + data_range * 0.5
else: # self.bound_type == 'mu_sigma'
lb = np.mean(data) - np.std(data) * self.sigma_multiplier
ub = np.mean(data) + np.std(data) * self.sigma_multiplier
if self.partitions:
old_partitions = self.partitions[:]
else:
old_partitions = pu.generate_t_partitions(self.nsets, lb, ub)
# 1) Compute the new partitions
self.partitions = pu.generate_t_partitions(self.nsets, lb, ub)
# 2) Verify the pertinence of the old sets centers with respect to the new partitions
old_centers = [p[1] for p in old_partitions]
f = fuzzify_x_list_t(old_centers, self.partitions)
# 3) Compute the final set of partitions
up_partitions = self.partitions + [
old_partitions[i] for i in range(len(f)) if f[i] < 0
]
up_partitions = sorted(up_partitions, key=lambda n_p: n_p[1])
self.partitions = up_partitions
self.nsets = len(self.partitions)
# 4) Compute the mappings required to update the rule base
map_old_new = fuzzify_x_list_t(old_centers, self.partitions)
# 5) Update rules
self.rule_base = rbm.update_rule_base(self.rule_base, map_old_new,
np.arange(len(self.partitions)),
self.order)
# 6) Add new rule
# Get pertinences
pertinence_list = pf.t_pertinence_list(
data[(len(data) - self.order - 1):], self.partitions)
# Build rule
rule = rbm.generate_rule(pertinence_list)
# Add rule to the rule base
rbm.add_rule(self.rule_base, rule, self.nsets, self.order)