def findParameters(trials = 100):
for runs in range(1,trials + 1):
net = rbf.rbf(train,traint,10,0,0)
net.rbftrain(train,traint,0.00001 + runs * 0.00001,2000)
output = net.rbffwd(alldata)
err = average(abs(output - alltgts))
print "Params: (" + str(0.00001 + runs * 0.00005) + ", 2000) / Error: " + str(err)
net = rbf.rbf(train,traint,10,0,0)
numruns = 2000 + runs * 10
net.rbftrain(train,traint,0.00001,numruns)
output = net.rbffwd(alldata)
err = average(abs(output - alltgts))
print "Params: (0.00001, " + str(numruns) + ") / Error: " + str(err)
def predict(train, traint, test, testt):
for res in range(0,2000):
tgts = []
for runs in range(0,20):
net = rbf.rbf(train,traint,16,0,0)
nbf = 9000
net.rbftrain(train,traint,0.00001,nbf)
output = net.rbffwd(alldata)
err = average(abs(output - alltgts))
tgts.append(str(14) + '\n' + str(1599) + str('\n'))
#f = open('res-' + str(runs) + '.csv','w')
#f.write("Error for this run: " + str(err) + "\n")
cnt = 0
for e in output:
#for i in range(1,20):
# f.write(str(e[0]) + str(','))
#f.write('\n')
tgts.append(str(average(e)) + str('\n'))
cnt += 1
#f.close()
f = open('tgts-' + str(err)[0:5] + '.csv','w')
for each in tgts:
f.write(each)
f.close()
def findRbfs(trials = 100):
for runs in range(1,trials + 1):
nrbfs = 1
net = rbf.rbf(train,traint,nrbfs + runs,0,0)
net.rbftrain(train,traint,0.00001,2000)
output = net.rbffwd(alldata)
err = average(abs(output - alltgts))
print "Num of RBFS: " + str(nrbfs + runs) + " / Error: " + str(err)
def setUp(self):
self.rbf_test1 = rbf.rbf(4,
0.5,
eta=0.25,
nIter=100,
thresh_type="logistic")
self.data = np.matrix([[0, 0], [0, 1], [1, 0], [1, 1]])
self.labels = np.array([0, 1, 1, 0])
self.rbf_test1.trainWeights(self.data, self.labels)
def setSurrogate(self,
type_surrogate='RBF',
min_snapshots=5,
max_snapshots=100,
max_degree=2,
initial_iteration=None,
no_keep=0,
expensive=0,
type_kernel=0,
type_solver=0):
self.min_snapshots = min_snapshots
self.max_snapshots = max_snapshots
if type_surrogate == 'SCM':
self.surrogate = scm(max_degree)
else:
self.surrogate = rbf(initial_iteration, no_keep, expensive,
type_kernel, type_solver)
def __init__(self, X, y, kernel, params = None, optim=False):
print("Initialising {} GP...".format(kernel))
if params is None:
params = {}
params['ln_noise'] = np.random.uniform(-1,1)
params['ln_signal'] = np.random.uniform(-1,1)
params['ln_length'] = np.random.uniform(-1,1)
if kernel.lower() == "rbf":
self.kernel = rbf(X,params)
elif kernel.lower() == "matern":
self.kernel = matern(X,params)
self.X = X
self.y = y
self.N = X.shape[0]
self.KMat(X,params)
if optim:
print("Optimising GP...")
self.optimize()
def CV_test(data, xs, ys, data_test, k):
ptxm = []
ptym = []
tptxm = []
tptym = []
for i in data:
x = 2 * (i[0] - xs.min()) / xs.ptp() - 1
y = 2 * (i[1] - ys.min()) / ys.ptp() - 1
ptx = [x, y]
ptxm.append(ptx)
label = i[2]
ptym.append(label)
for i in data_test:
x = 2 * (i[0] - xs.min()) / xs.ptp() - 1
y = 2 * (i[1] - ys.min()) / ys.ptp() - 1
ptx = [x, y]
tptxm.append(ptx)
label = i[2]
tptym.append(label)
nn = rbf(k)
cv_error = 0.0
in_error = 0.0
for i in range(len(data)):
_ptxm = list(ptxm)
_ptym = list(ptym)
test_x = np.array([_ptxm.pop(i)])
test_y = _ptym.pop(i)
x_matrix = np.array(_ptxm)
y_matrix = np.array(np.transpose(_ptym))
nn.train(x_matrix, y_matrix)
if nn.predict(test_x) * test_y < 0:
cv_error += 1
cv_error /= len(data)
ptxm = np.array(ptxm)
ptym = np.array(ptym)
nn.train(ptxm, ptym)
tptxm = np.array(tptxm)
tptym = np.array(tptym)
y_test = nn.predict(tptxm)
test_error = (np.multiply(y_test, tptym) < 0).sum() / float(len(tptym))
y_test = nn.predict(ptxm)
in_error = (np.multiply(y_test, ptym) < 0).sum() / float(len(tptym))
return test_error, cv_error
def predictKfold(train, traint, test, testt, runs, datafile):
err = []
cnt = 0
pred = []
tgts = []
allpred = []
alltgt = []
# per datafile params
if (datafile[0] == 'w'):
rowcnt = '4898'
outfn = 'ww-'
else:
rowcnt = '1599'
outfn = 'rw-'
for res in range(0,runs):
tgts.append(str(13) + '\n' + str(rowcnt) + str('\n'))
pred.append(str(14) + '\n' + str(rowcnt) + str('\n'))
for folds in range(0,5):
net = rbf.rbf(train[folds],traint[folds],22,0,0)
nbf = 2000
net.rbftrain(train[folds],traint[folds],0.00001,nbf)
output = net.rbffwd(test[folds])
err.append(average(abs(output - testt[folds])))
for e in output:
pred.append(str(e[0]) + str('\n'))
allpred.append(e[0])
for e in testt[folds]:
tgts.append(str(e[0])[0] + str('\n'))
alltgt.append (int(str(e[0])[0]))
print(average(err))
f = open(outfn + 'pred-' + str(average(err))[0:5] + '.csv','w')
for each in pred:
f.write(each)
f.close()
f = open(outfn + 'tgts-' + str(average(err))[0:5] + '.csv','w')
for each in tgts:
f.write(each)
f.close()
def train(k, xx, yy, data):
ptxm = []
ptym = []
for i in data:
ptx = [i[0], i[1]]
ptxm.append(ptx)
label = i[2]
ptym.append(label)
nn = rbf(k)
ptxm = np.array(ptxm)
ptym = np.array(ptym)
#print(ptxm)
#print(ptym)
nn.train(ptxm, ptym)
zz = []
#print(nn.predict(ptxm))
x_len, y_len = np.shape(xx)
tptx = []
for i in range(x_len):
for j in range(y_len):
px = xx[i][j]
py = yy[i][j]
ptx = [px, py]
tptx.append(ptx)
#print(tptx)
z = nn.predict(np.array(tptx))
#print(z)
for i in range(x_len):
z0 = []
for j in range(y_len):
x = z[i * x_len + j]
if x < 0:
z0.append(-1)
else:
z0.append(1)
zz.append(z0)
zz = np.array(zz)
return zz
def rBasis():
pima = np.loadtxt(
'C:\Users\subha\PycharmProjects\Radial_Basis/pima-indians-diabetes.data',
delimiter=',')
pima[:, :8] = pima[:, :8] - pima[:, :8].mean(axis=0)
imax = np.concatenate((pima.max(axis=0) * np.ones(
(1, 9)), pima.min(axis=0) * np.ones((1, 9))),
axis=0).max(axis=0)
pima[:, :8] = pima[:, :8] / imax[:8]
target = np.zeros((np.shape(pima)[0], 3))
indices = np.where(pima[:, 8] == 0)
target[indices, 0] = 1
indices = np.where(pima[:, 8] == 1)
target[indices, 1] = 1
indices = np.where(pima[:, 8] == 2)
target[indices, 2] = 1
order = range(np.shape(pima)[0])
np.random.shuffle(order)
pima = pima[order, :]
target = target[order, :]
train = pima[::2, :8]
traint = target[::2]
valid = pima[3::4, 0:8]
validt = target[3::4]
test = pima[1::4, 0:8]
testt = target[1::4]
net = rbf.rbf(train, traint, 3)
net.rbftrain(train, traint, 0.25, 2000)
cm_train, accuracy_train = net.confmat(train, traint)
cm_test, accuracy_test = net.confmat(test, testt)
return cm_train, accuracy_train, cm_test, accuracy_test
if i > 0:
set_size[i] = set_size[i] + set_size[i - 1]
train = data[:set_size[0], 0:col - output_size]
traint = target[:set_size[0]]
valid = data[set_size[0]:set_size[1], 0:col - output_size]
validt = target[set_size[0]:set_size[1]]
test = data[set_size[1]:set_size[2], 0:col - output_size]
testt = target[set_size[1]:set_size[2]]
# Train the network
# MLP
if run[0]:
import mlp as mlp_module
mlp = mlp_module.mlp(train, traint, nhidden, outtype='softmax')
mlp.earlystopping(train, traint, valid, validt, eta)
mlp_correct_pct = mlp.confmat(test, testt)
if save_results:
mlp.get_data(resultPath, filename, SEED)
print 'MLP weights saved'
#RBF
if run[1]:
import rbf as rbf_module
rbf = rbf_module.rbf(train, traint, nRBF, 1, 1)
rbf.rbftrain(train, traint, 0.25, 2000)
rbf_correct_pct = rbf.confmat(test, testt)
if save_results:
rbf.get_data(resultPath + filename)
print 'RBF weights saved'
import numpy as np import rbf iris = load_iris() X = iris.data y = iris.target n_rbf = 5 sigma = 1 use_kmeans = 1 eta = 0.25 n_iterations = 2000 order = range(X.shape[0]) np.random.shuffle(order) X = X[order] y = y[order] lb = LabelBinarizer() y = lb.fit_transform(y) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25) net = rbf.rbf(X_train, y_train, n_rbf, sigma, use_kmeans) net.rbf_train(X_train, y_train, eta, n_iterations) net.confmat(X_test, y_test)
target = np.zeros((np.shape(iris)[0], 3)); indices = np.where(iris[:, 4] == 0) target[indices, 0] = 1 indices = np.where(iris[:, 4] == 1) target[indices, 1] = 1 indices = np.where(iris[:, 4] == 2) target[indices, 2] = 1 order = list(range(np.shape(iris)[0])) np.random.shuffle(order) iris = iris[order, :] target = target[order, :] train = iris[::2, 0:4] traint = target[::2] valid = iris[1::4, 0:4] validt = target[1::4] test = iris[3::4, 0:4] testt = target[3::4] # print train.max(axis=0), train.min(axis=0) import rbf net = rbf.rbf(train, traint, 5, 1, 1) net.rbftrain(train, traint, 0.25, 2000) # net.confmat(train,traint) net.confmat(test, testt)
import os
import string
import pylab as pl
import numpy as np
#find the data files in the source directory
os.chdir(os.path.abspath(os.path.dirname(__file__)))
train_data = np.genfromtxt('datatraining.txt', delimiter=',', skip_header=1)
test_data = np.genfromtxt('datatest.txt', delimiter=',', skip_header=1)
test_data2 = np.genfromtxt('datatest2.txt', delimiter=',', skip_header=1)
#splitting of the data into training and testing matrices
trainin = train_data[:1333:, 2:7:]
traintgt = train_data[:1333:, 7::]
testin = test_data[::2, 2:7:]
testtgt = test_data[::2, 7::]
test2in = test_data2[:1333:, 2:7:]
test2tgt = test_data2[:1333:, 7::]
#multilayer perceptron model
print('Multilayer Perceptron Model')
net = mlp.mlp(trainin, traintgt, 2, outtype='linear')
net.earlystopping(trainin, traintgt, testin, testtgt, 0.001, 1000)
net.confmat(test2in, test2tgt)
#rbf model
print('RBF Model')
net = rbf.rbf(trainin, traintgt, 6, 0, 1)
net.rbftrain(trainin, traintgt, 0.01, 2000)
net.confmat(testin, testtgt)
comm_world.Recv(received_data, source=group_leader_ids[0], tag=1)
Model.SetNoisyObservation(received_data[no_parameters + 1:])
print('artificial_observation_without_noise:', received_data[1:])
print('artificial observation with noise:', Model.observation)
SF = SamplingFramework.SamplingFramework(no_chains, stages, max_snapshots,
min_snapshots, no_solvers)
MHsampler.MHsampler.SF = SF
MHsampler.MHsampler.seed = seed
initial_samples = lhs_norm.lhs_norm(Model, SF.no_chains, seed)
print("initial samples:", initial_samples)
G_data = np.zeros([0, no_observations + no_parameters])
if surr_type == 'SCM':
Surrogate = scm.scm()
else:
Surrogate = rbf.rbf()
Surrogate.parameters = surrogate_parameters
# G_data = np.vstack([G_data,np.append(artificial_observation_without_noise,artificial_real_parameters)]) # should not be known
# SF.shared_queue_surrogate.put([artificial_real_parameters,artificial_observation_without_noise]) # should not be known
if __name__ == '__main__':
jobs = []
for i in range(SF.no_chains + SF.no_helpers):
p = mp.Process(target=process_worker.worker,
args=(i, comm_world, Model, SF, Surrogate,
no_solvers, initial_samples, group_leader_ids,
proposalStd))
p.start()
jobs.append(p)
process_master.master(comm_world, Model, SF, G_data, no_solvers,
group_leader_ids)
if __name__ == '__main__':
#target[indices,1] = 1 target = zeros((shape(iris)[0], 3)) indices = where(iris[:, 4] == 0) target[indices, 0] = 1 indices = where(iris[:, 4] == 1) target[indices, 1] = 1 indices = where(iris[:, 4] == 2) target[indices, 2] = 1 order = range(shape(iris)[0]) random.shuffle(order) iris = iris[order, :] target = target[order, :] train = iris[::2, 0:4] traint = target[::2] valid = iris[1::4, 0:4] validt = target[1::4] test = iris[3::4, 0:4] testt = target[3::4] #print train.max(axis=0), train.min(axis=0) import rbf net = rbf.rbf(train, traint, 5, 1, 1) net.rbftrain(train, traint, 0.25, 2000) #net.confmat(train,traint) net.confmat(test, testt)
])
class1_Y = np.tile(np.array([1, 0, 0]), (class1_n_points, 1))
class2_Y = np.tile(np.array([0, 1, 0]), (class2_n_points, 1))
class3_Y = np.tile(np.array([0, 0, 1]), (class3_n_points, 1))
n_samples = class1_n_points + class2_n_points + class3_n_points
permutation = np.random.permutation(n_samples)
X = np.take(np.vstack([class1_X, class2_X, class3_X]), permutation, axis=0)
Y = np.take(np.vstack([class1_Y, class2_Y, class3_Y]), permutation, axis=0)
x_train, y_train, x_test, y_test, x_valid, y_valid = split_dataset(
X, Y, 0.2, 0)
classifier = rbf()
spread = 0.3
classifier.fit(x_train, y_train, spread=spread)
y_pred = np.where(classifier.predict(x_train) > 0.5, 1, 0)
correctly_classified = row_wise_equal(y_pred, y_train)
print(
f'{correctly_classified}/{x_train.shape[0]} ({correctly_classified / x_train.shape[0]}) points of train set are classified correctly!'
)
y_pred = np.where(classifier.predict(x_test) > 0.5, 1, 0)
correctly_classified = row_wise_equal(y_pred, y_test)
print(
f'{correctly_classified}/{x_test.shape[0]} ({correctly_classified / x_test.shape[0]}) points of test set are classified correctly!'
)
""" #logical OR inputs = np.array([[1,1],[1,0],[0,1],[0,0]]) targets = np.array([[1],[1],[1],[0]]) """ #logical XOR inputs = np.array([[1,1],[1,0],[0,1],[0,0]]) targets = np.array([[0],[1],[1],[0]]) """ #identity matrix inputs = np.array([[1,1],[1,0],[0,1],[0,0]]) targets = np.array([[1,1],[1,0],[0,1],[0,0]]) """ #use one-layer perceptron p = pcn.pcn(inputs,targets,0.2,'linear','batch') p.pcntrain(10000) p.confmat(inputs,targets) #use rbf network p = rbf.rbf(inputs,targets,4,0,1,0.2,'linear','batch') p.rbftrain(10000) p.confmat(inputs,targets) #use two-layer perceptron p = mlpcn.mlpcn(inputs,targets,4,0.2,'linear','batch') p.mlptrain(10000) p.confmat(inputs,targets)
if i > 0:
set_size[i] = set_size[i] + set_size[i - 1]
train = data[:set_size[0],0:col-output_size]
traint = target[:set_size[0]]
valid = data[set_size[0]:set_size[1],0:col-output_size]
validt = target[set_size[0]:set_size[1]]
test = data[set_size[1]:set_size[2],0:col-output_size]
testt = target[set_size[1]:set_size[2]]
# Train the network
# MLP
if run[0]:
import mlp as mlp_module
mlp = mlp_module.mlp(train,traint,nhidden,outtype='softmax')
mlp.earlystopping(train,traint,valid,validt,eta)
mlp_correct_pct = mlp.confmat(test,testt)
if save_results:
mlp.get_data(resultPath, filename, SEED)
print 'MLP weights saved'
#RBF
if run[1]:
import rbf as rbf_module
rbf = rbf_module.rbf(train,traint,nRBF,1,1)
rbf.rbftrain(train,traint,0.25,2000)
rbf_correct_pct = rbf.confmat(test,testt)
if save_results:
rbf.get_data(resultPath + filename)
print 'RBF weights saved'
def lspi():
gridworld = Gridworld()
gamma = .8
beta = 1
numEpisodes = 10
beta_factor = 1 - 5 * numEpisodes / 10000
# Sample the state space with different grid sizes
X_1, X_2 = np.meshgrid(np.arange(.5, 7, 1),
np.arange(.5, 7, 1),
indexing='ij')
X_1m, X_2m = np.meshgrid(np.arange(.5, 7, 0.5),
np.arange(.5, 7, 0.5),
indexing='ij')
# initialize the policy randomly. should give for each state in s (nx2) a
# random action of the form (r,a), where r ~ Unif[0,1] and a ~ Unif[0,2pi].
# pi = lambda s: ...
# samples from initial distribution n starting positions (you can start
# with a random initialization in the entire gridworld)
# initialDistribution = lambda n: ...
converged = False
# generate an ndgrid over the state space for the centers of the basis
# functions
X1, X2, A1, A2 = np.meshgrid(np.arange(.5, 7, 1), np.arange(.5, 7, 1),
np.arange(-1, 2), np.arange(-1, 2))
# NOTE: the policy returns the action in polar coordinates while the basis
# functions use cartesian coordinates!!! You have to convert between these
# representations.
# matrix of the centers
c = np.column_stack(
(np.transpose(X1.flatten()), np.transpose(X2.flatten()),
np.transpose(A1.flatten()), np.transpose(A2.flatten())))
# number of basis functions
# k = ...
# initialize weights
# w = ...
# compute bandwiths with median trick
bw = np.zeros(4)
for i in range(4):
dist = pdist(c[:, [i]])
bw[i] = np.sqrt(np.median(dist**2)) * .4
# feature function (making use of rbf)
# feature = lambda x_: ...
# time step
t = 0
# initialize A and b
# A = ...
# b = ...
while not converged:
# Policy evaluation
# sample data
s1, a, r, s2 = sampleData(gridworld, pi, initialDistribution,
numEpisodes, 50)
# compute actions in cartesian space
ac1, ac2 = pol2cart(a[:, 0, np.newaxis], a[:, 1, np.newaxis])
# compute PHI
# PHI = ...
# compute PPI
# PPI = ...
# update A and b
# A = ...
# b = ...
# compute new w
w_old = w
# w = ...
# Policy improvement
# pi = ...
beta = beta_factor * beta
t = t + 1
# Check for convergence
if np.abs(w - w_old).sum() / len(w) < 0.05:
converged = True
print(t, ' - ', beta, ' - ', np.abs(w - w_old).sum() / len(w))
### plotting
a = policy(np.hstack((X_1m.reshape(-1, 1), X_2m.reshape(-1, 1))),
feature, w, 0)
ax1, ax2 = pol2cart(a[:, 0].reshape(-1, 1), a[:, 1].reshape(-1, 1))
phi = rbf(
np.hstack((X_1m.reshape(-1, 1), X_2m.reshape(-1, 1), ax1, ax2)), c,
bw)
Q = phi.dot(w)
n_plot = len(X_1m)
plot_a = np.hstack((ax1, ax2)).reshape((n_plot, n_plot, 2))
plot_V = Q.reshape((n_plot, n_plot))
plotPiV(plot_a, plot_V, vmin=-5, vmax=5, block=False)
plotPiV(plot_a, plot_V, vmin=-5, vmax=5)
test = iris[3::4, 0:4]
testt = target[3::4]
"""
#use one-layer perceptron
net = pcn.pcn(train,traint,0.2,'softmax','batch')
print('\nperceptron training...')
(trainerror,validerror) = net.pcntrain_automatic(valid,validt,100)
print('Training stop after',len(trainerror)*100,'iterations')
print('Final Train Error',net.errfunc(net.pcnfwd(train,True),traint))
print('Final Valid Error',net.errfunc(net.pcnfwd(valid,True),validt))
"""
#use rbf perceptron
net = rbf.rbf(train, traint, 10, 0, 1, 0.2, 'softmax', 'batch')
print('\nperceptron training ...')
(trainerror, validerror) = net.rbftrain_automatic(valid, validt, 100)
print('Training stop after', len(trainerror) * 100, 'iterations')
print('Final Train Error', net.errfunc(net.rbffwd(train, 2), traint))
print('Final Valid Error', net.errfunc(net.rbffwd(valid, 2), validt))
"""
#use two-layer perceptron
net = mlpcn.mlpcn(train,traint,4,0.2,'softmax','batch')
print('perceptron training...')
(trainerror,validerror) = net.mlptrain_automatic(train,traint,valid,validt,100)
print('Training stop after',len(trainerror)*100,'iterations')
def execute(self):
[rbf(expression) for expression in self._params_to_execute]
def rbnn(x_train, y_train, theta, funtol):
# Initialization
q = (np.shape(x_train))[0]
r = (np.shape(x_train))[1]
x = x_train.copy() # Create a copy of Training Data
p = np.zeros([q, q]) # Design Matrix
err = np.zeros([q, 1]) # Store Errors associated with columns of p
c = np.zeros([q, r]) # Store neurons centers
phi_col = np.zeros([q, 1]) # Array to update design matrix for actual error calculations
showiterinfo = True # Display iterations
# Create Lists for Plots
x_data = [] # Epochs
y_data = [] # MSE
# Create Design Matrix
for i in range(q):
for j in range(q):
p[j][i] = rbf(x[j:j + 1, 0:r], x[i:i + 1, 0:r], theta) # RBF function call
# Find column vector in p with maximum error
for k in range(q):
a = p[0:q, k:k + 1]
g = np.dot(np.transpose(a), y_train) / np.dot(np.transpose(a), a)
err[k] = pow(g, 2) * np.dot(np.transpose(a), a) / np.dot(np.transpose(y_train), y_train)
maxerr = err.max()
j = err.argmax()
wj = p[0:q, j:j + 1]
p = np.delete(p, j, 1)
err = np.delete(err, j, 0)
c[0:1, 0:r] = x[j:j + 1, 0:r]
x = np.delete(x, j, 0)
# Calculate mean-squared error of the network
for i in range(q):
# Update Design Matrix for Training Data
phi_col[i][0] = rbf(x_train[i:i + 1, 0:r], c[0:1, 0:r], theta)
bias = np.ones([q, 1]) # Add bias term
phi = np.concatenate((bias, phi_col), axis=1)
lw = np.dot(np.linalg.pinv(phi[0:q, 0:2]), y_train)
mse = np.mean(np.square(np.dot(phi[0:q, 0:2], lw) - y_train))
x_data.append(1)
y_data.append(mse)
# Display iteration info
if showiterinfo:
print("Epoch: " + str(1) + " MSE: " + str(mse))
# Main loop
for it in range(q - 1):
alpha = np.dot(np.transpose(wj), p) / np.dot(np.transpose(wj), wj)
p = p - np.dot(wj, alpha)
e = (np.shape(p))[1]
# Find column vector in p with maximum error
for k in range(e):
a = p[0:q, k:k + 1]
g = np.dot(np.transpose(a), y_train) / np.dot(np.transpose(a), a)
err[k] = pow(g, 2) * np.dot(np.transpose(a), a) / np.dot(np.transpose(y_train), y_train)
maxerr = err.max()
j = err.argmax()
wj = p[0:q, j:j + 1]
p = np.delete(p, j, 1)
err = np.delete(err, j, 0)
c[it + 1:it + 2, 0:r] = x[j:j + 1, 0:r]
x = np.delete(x, j, 0)
# Calculate mean-squared error of the network
for i in range(q):
phi_col[i][0] = rbf(x_train[i:i + 1, 0:q - 1], c[it + 1:it + 2, 0:r], theta)
# Update Design Matrix for Training Data
phi = np.concatenate((phi, phi_col), axis=1)
lw = np.dot(np.linalg.pinv(phi[0:q, 0:it + 3]), y_train)
mse = np.mean(np.square(np.dot(phi[0:q, 0:it + 3], lw) - y_train))
x_data.append(it + 2)
y_data.append(mse)
# Display iteration info
if showiterinfo:
print("Epoch: " + str(it + 2) + " MSE: " + str(mse))
# Check for convergence
if mse <= funtol:
print("Algorithm Stopped:Mean-squared error less than specified tolerance")
for j in range(q-1, it+1, -1):
c = np.delete(c, j, 0)
break
if it == (q - 2):
print("Algorithm Stopped: Maximum number of neurons reached")
break
# Plot for Best and Average Cost Function with Iterations
plt.plot(x_data, y_data, '-b')
plt.xlabel('Epochs')
plt.ylabel("MSE")
plt.grid()
plt.legend(["Training"], loc="upper right", frameon=False)
plt.show()
return c, lw