def plot(self, y="return", x="learning_steps", save=False):
"""Plots the performance of the experiment
This function has only limited capabilities.
For more advanced plotting of results consider
:py:class:`Tools.Merger.Merger`.
"""
labels = problems.rlpy.Tools.results.default_labels
performance_fig = plt.figure("Performance")
res = self.result
plt.plot(res[x], res[y], '-bo', lw=3, markersize=10)
plt.xlim(0, res[x][-1] * 1.01)
y_arr = np.array(res[y])
m = y_arr.min()
M = y_arr.max()
delta = M - m
if delta > 0:
plt.ylim(m - .1 * delta - .1, M + .1 * delta + .1)
xlabel = labels[x] if x in labels else x
ylabel = labels[y] if y in labels else y
plt.xlabel(xlabel, fontsize=16)
plt.ylabel(ylabel, fontsize=16)
if save:
path = os.path.join(self.full_path,
"{:3}-performance.pdf".format(self.exp_id))
performance_fig.savefig(path, transparent=True, pad_inches=.1)
plt.ioff()
plt.show()
def batchDiscover(self, td_errors, phi, states):
# Discovers features using iFDD in batch setting.
# TD_Error: p-by-1 (How much error observed for each sample)
# phi: n-by-p features corresponding to all samples (each column corresponds to one sample)
# self.batchThreshold is the minimum relevance value for the feature to
# be expanded
SHOW_PLOT = 0 # Shows the histogram of relevances
maxDiscovery = self.maxBatchDiscovery
n = self.features_num # number of features
p = len(td_errors) # Number of samples
counts = np.zeros((n, n))
relevances = np.zeros((n, n))
for i in range(p):
phiphiT = np.outer(phi[i, :], phi[i, :])
if self.iFDDPlus:
relevances += phiphiT * td_errors[i]
else:
relevances += phiphiT * abs(td_errors[i])
counts += phiphiT
# Remove Diagonal and upper part of the relevances as they are useless
relevances = np.triu(relevances, 1)
non_zero_index = np.nonzero(relevances)
if self.iFDDPlus:
# Calculate relevances based on theoretical results of ICML 2013
# potential submission
relevances[non_zero_index] = np.divide(
np.abs(relevances[non_zero_index]),
np.sqrt(counts[non_zero_index]))
else:
# Based on Geramifard11_ICML Paper
relevances[non_zero_index] = relevances[non_zero_index]
# Find indexes to non-zero excited pairs
# F1 and F2 are the parents of the potentials
(F1, F2) = relevances.nonzero()
relevances = relevances[F1, F2]
if len(relevances) == 0:
# No feature to add
self.logger.debug("iFDD Batch: Max Relevance = 0")
return False
if SHOW_PLOT:
e_vec = relevances.flatten()
e_vec = e_vec[e_vec != 0]
e_vec = np.sort(e_vec)
plt.ioff()
plt.plot(e_vec, linewidth=3)
plt.show()
# Sort based on relevances
# We want high to low hence the reverse: [::-1]
sortedIndices = np.argsort(relevances)[::-1]
max_relevance = relevances[sortedIndices[0]]
# Add top <maxDiscovery> features
self.logger.debug(
"iFDD Batch: Max Relevance = {0:g}".format(max_relevance))
added_feature = False
new_features = 0
for j in range(len(relevances)):
if new_features >= maxDiscovery:
break
max_index = sortedIndices[j]
f1 = F1[max_index]
f2 = F2[max_index]
relevance = relevances[max_index]
if relevance > self.batchThreshold:
# print "Inspecting",
# f1,f2,'=>',self.getStrFeatureSet(f1),self.getStrFeatureSet(f2)
if self.inspectPair(f1, f2, np.inf):
self.logger.debug(
'New Feature %d: %s, Relevance = %0.3f' %
(self.features_num - 1,
self.getStrFeatureSet(self.features_num - 1),
relevances[max_index]))
new_features += 1
added_feature = True
else:
# Because the list is sorted, there is no use to look at the
# others
break
return (
# A signal to see if the representation has been expanded or not
added_feature)
def showDomain(self, a):
s = self.state
# Plot the car and an arrow indicating the direction of accelaration
# Parts of this code was adopted from Jose Antonio Martin H.
# <*****@*****.**> online source code
pos, vel = s
if self.domain_fig is None: # Need to initialize the figure
self.domain_fig = plt.figure("Mountain Car Domain")
# plot mountain
mountain_x = np.linspace(self.XMIN, self.XMAX, 1000)
mountain_y = np.sin(3 * mountain_x)
plt.gca(
).fill_between(mountain_x,
min(mountain_y) - self.CAR_HEIGHT * 2,
mountain_y,
color='g')
plt.xlim([self.XMIN - .2, self.XMAX])
plt.ylim(
[min(mountain_y) - self.CAR_HEIGHT * 2,
max(mountain_y) + self.CAR_HEIGHT * 2])
# plot car
self.car = lines.Line2D([], [], linewidth=20, color='b', alpha=.8)
plt.gca().add_line(self.car)
# Goal
plt.plot(self.GOAL, np.sin(3 * self.GOAL), 'yd', markersize=10.0)
plt.axis('off')
plt.gca().set_aspect('1')
self.domain_fig = plt.figure("Mountain Car Domain")
#pos = 0
#a = 0
car_middle_x = pos
car_middle_y = np.sin(3 * pos)
slope = np.arctan(3 * np.cos(3 * pos))
car_back_x = car_middle_x - self.CAR_WIDTH * np.cos(slope) / 2.
car_front_x = car_middle_x + self.CAR_WIDTH * np.cos(slope) / 2.
car_back_y = car_middle_y - self.CAR_WIDTH * np.sin(slope) / 2.
car_front_y = car_middle_y + self.CAR_WIDTH * np.sin(slope) / 2.
self.car.set_data([car_back_x, car_front_x], [car_back_y, car_front_y])
# wheels
# plott(x(1)-0.05,sin(3*(x(1)-0.05))+0.06,'ok','markersize',12,'MarkerFaceColor',[.5 .5 .5]);
# plot(x(1)+0.05,sin(3*(x(1)+0.05))+0.06,'ok','markersize',12,'MarkerFaceColor',[.5 .5 .5]);
# Arrows
if self.actionArrow is not None:
self.actionArrow.remove()
self.actionArrow = None
if self.actions[a] > 0:
self.actionArrow = fromAtoB(
car_front_x, car_front_y,
car_front_x + self.ARROW_LENGTH *
np.cos(slope), car_front_y +
self.ARROW_LENGTH * np.sin(slope),
#car_front_x + self.CAR_WIDTH*cos(slope)/2., car_front_y + self.CAR_WIDTH*sin(slope)/2.+self.CAR_HEIGHT,
'k', "arc3,rad=0",
0, 0, 'simple'
)
if self.actions[a] < 0:
self.actionArrow = fromAtoB(
car_back_x, car_back_y,
car_back_x - self.ARROW_LENGTH *
np.cos(slope), car_back_y -
self.ARROW_LENGTH * np.sin(slope),
#car_front_x + self.CAR_WIDTH*cos(slope)/2., car_front_y + self.CAR_WIDTH*sin(slope)/2.+self.CAR_HEIGHT,
'r', "arc3,rad=0",
0, 0, 'simple'
)
plt.draw()
def batchDiscover(self, td_errors, phi, states):
"""
:param td_errors: p-by-1 vector, error associated with each state
:param phi: p-by-n matrix, vector-valued feature function evaluated at
each state.
:param states: p-by-(statedimension) matrix, each state under test.
Discovers features using OMPTD
1. Find the index of remaining features in the bag \n
2. Calculate the inner product of each feature with the TD_Error vector \n
3. Add the top maxBatchDiscovery features to the selected features \n
OUTPUT: Boolean indicating expansion of features
"""
if len(self.remainingFeatures) == 0:
# No More features to Expand
return False
SHOW_RELEVANCES = 0 # Plot the relevances
self.calculateFullPhiNormalized(states)
relevances = np.zeros(len(self.remainingFeatures))
for i, f in enumerate(self.remainingFeatures):
phi_f = self.fullphi[:, f]
relevances[i] = np.abs(np.dot(phi_f, td_errors))
if SHOW_RELEVANCES:
e_vec = relevances.flatten()
e_vec = e_vec[e_vec != 0]
e_vec = np.sort(e_vec)
plt.plot(e_vec, linewidth=3)
plt.ioff()
plt.show()
plt.ion()
# Sort based on relevances
# We want high to low hence the reverse: [::-1]
sortedIndices = np.argsort(relevances)[::-1]
max_relevance = relevances[sortedIndices[0]]
# Add top <maxDiscovery> features
self.logger.debug("OMPTD Batch: Max Relevance = %0.3f" % max_relevance)
added_feature = False
to_be_deleted = [] # Record the indices of items to be removed
for j in range(min(self.maxBatchDiscovery, len(relevances))):
max_index = sortedIndices[j]
f = self.remainingFeatures[max_index]
relevance = relevances[max_index]
# print "Inspecting %s" % str(list(self.iFDD.getFeature(f).f_set))
if relevance >= self.batchThreshold:
self.logger.debug(
'New Feature %d: %s, Relevance = %0.3f' %
(self.features_num,
str(np.sort(list(self.iFDD.getFeature(f).f_set))),
relevances[max_index]))
to_be_deleted.append(max_index)
self.selectedFeatures.append(f)
self.features_num += 1
added_feature = True
else:
# Because the list is sorted, there is no use to look at the
# others
break
self.remainingFeatures = np.delete(self.remainingFeatures,
to_be_deleted)
return added_feature