def _plot_impl(self,
y="return",
x="learning_steps",
save=False,
show=True):
labels = rlpy.tools.results.default_labels
performance_fig = plt.figure("Performance")
res = self.result
plt.plot(res[x], res[y], lw=2, markersize=4, marker=MARKERS[0])
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 - 0.1 * delta - 0.1, M + 0.1 * delta + 0.1)
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,
"{:03}-performance.pdf".format(self.exp_id))
performance_fig.savefig(path, transparent=True, pad_inches=0.1)
if show:
plt.ioff()
plt.show()
def show_domain(self, a=0):
s = self.state
# Draw the environment
if self.domain_fig is None:
self.move_fig = plt.subplot(111)
s = s.reshape((self.BOARD_SIZE, self.BOARD_SIZE))
self.domain_fig = plt.imshow(s,
cmap="FlipBoard",
interpolation="nearest",
vmin=0,
vmax=1)
plt.xticks(np.arange(self.BOARD_SIZE), fontsize=FONTSIZE)
plt.yticks(np.arange(self.BOARD_SIZE), fontsize=FONTSIZE)
# pl.tight_layout()
a_row, a_col = id2vec(a, [self.BOARD_SIZE, self.BOARD_SIZE])
self.move_fig = self.move_fig.plot(a_col,
a_row,
"kx",
markersize=30.0)
plt.show()
a_row, a_col = id2vec(a, [self.BOARD_SIZE, self.BOARD_SIZE])
self.move_fig.pop(0).remove()
# print a_row,a_col
# Instead of '>' you can use 'D', 'o'
self.move_fig = plt.plot(a_col, a_row, "kx", markersize=30.0)
s = s.reshape((self.BOARD_SIZE, self.BOARD_SIZE))
self.domain_fig.set_data(s)
plt.draw()
def show_domain(self, a):
s = self.state
# Draw the environment
fig = plt.figure("IntruderMonitoring")
if self.domain_fig is None:
self.domain_fig = plt.imshow(
self.map,
cmap="IntruderMonitoring",
interpolation="nearest",
vmin=0,
vmax=3,
)
plt.xticks(np.arange(self.COLS), fontsize=FONTSIZE)
plt.yticks(np.arange(self.ROWS), fontsize=FONTSIZE)
plt.show()
if self.ally_fig is not None:
self.ally_fig.pop(0).remove()
self.intruder_fig.pop(0).remove()
s_ally = s[0:self.NUMBER_OF_AGENTS * 2].reshape((-1, 2))
s_intruder = s[self.NUMBER_OF_AGENTS * 2:].reshape((-1, 2))
self.ally_fig = plt.plot(
s_ally[:, 1],
s_ally[:, 0],
"bo",
markersize=30.0,
alpha=0.7,
markeredgecolor="k",
markeredgewidth=2,
)
self.intruder_fig = plt.plot(
s_intruder[:, 1],
s_intruder[:, 0],
"g>",
color="gray",
markersize=30.0,
alpha=0.7,
markeredgecolor="k",
markeredgewidth=2,
)
fig.canvas.draw()
fig.canvas.flush_events()
def show_domain(self, a=None):
if a is not None:
a = self.actions[a]
T = np.empty((self.d, 2))
T[:, 0] = np.cos(self.theta)
T[:, 1] = np.sin(self.theta)
R = np.dot(self.P, T)
R1 = R - 0.5 * self.lengths[:, None] * T
R2 = R + 0.5 * self.lengths[:, None] * T
Rx = np.hstack([R1[:, 0], R2[:, 0]]) + self.pos_cm[0]
Ry = np.hstack([R1[:, 1], R2[:, 1]]) + self.pos_cm[1]
fig = plt.figure("Swimmer")
if self.swimmer_lines is None:
plt.plot(0.0, 0.0, "ro")
self.swimmer_lines = plt.plot(Rx, Ry)[0]
self.action_text = plt.text(-2, -8, str(a))
plt.xlim(-5, 15)
plt.ylim(-10, 10)
else:
self.swimmer_lines.set_data(Rx, Ry)
self.action_text.set_text(str(a))
fig.canvas.draw()
fig.canvas.flush_events()
def batch_discover(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 max_batch_discovery features to the selected features \n
OUTPUT: Boolean indicating expansion of features
"""
if len(self.remaining_features) == 0:
# No More features to Expand
return False
self.calculate_full_phi_normalized(states)
relevances = np.zeros(len(self.remaining_features))
for i, f in enumerate(self.remaining_features):
phi_f = self.fullphi[:, f]
relevances[i] = np.abs(np.dot(phi_f, td_errors))
if self.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.max_batch_discovery, len(relevances))):
max_index = sortedIndices[j]
f = self.remaining_features[max_index]
relevance = relevances[max_index]
# print "Inspecting %s" % str(list(self.iFDD.getFeature(f).f_set))
if relevance >= self.batch_threshold:
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.selected_features.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.remaining_features = np.delete(self.remaining_features,
to_be_deleted)
return added_feature
def batch_discover(self, td_errors, phi, states):
"""
Discovers features using iFDD in batch setting.
self.batch_threshold is the minimum relevance value for the feature
to be expanded.
:param td_errors: p-by-1 (How much error observed for each sample)
:param phi: n-by-p features corresponding to all samples
(each column corresponds to one sample).
"""
maxDiscovery = self.max_batch_discovery
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 self.debug:
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.batch_threshold:
# 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
# A signal to see if the representation has been expanded or not
return added_feature