def mrmr_selection(input_data_x, input_data_y):
max_mi = np.argmax(mutual_info_classif(input_data_x, input_data_y))
input_data_y = np.reshape(input_data_y, (-1, 1))
select_index = []
select_index.append(max_mi)
selected_features = input_data_x[:, max_mi].reshape(-1, 1)
diff = []
for x in range(0, num_of_features):
#print(str(x)+' / '+str(num_of_features))
for i in range(0, input_data_x.shape[1]):
key = 0
for j in range(0, len(select_index)):
if i == select_index[j]:
key = 1
if key == 0:
feature = input_data_x[:, i]
MI = 0
count = 0
for k in range(0, selected_features.shape[1]):
MI = MI + mutinf(
100, feature.reshape(-1, 1),
np.reshape(selected_features[:, k], (-1, 1)))
count = count + 1
diff.append(
mutinf(100, feature.reshape(-1, 1), input_data_y) -
float(MI) / count)
else:
diff.append(-99.9)
max_new = np.argmax(diff)
select_index.append(max_new)
del diff[:]
selected_features = np.append(selected_features,
input_data_x[:, max_new].reshape(-1, 1),
axis=1)
return select_index
def bias_value(input_data_x, input_data_y):
bias_num = []
input_data_y = np.reshape(input_data_y, (-1, 1))
input_data_x = np.array(input_data_x)
input_data_x_old = input_data_x
high_score = mutinf(100, input_data_x, input_data_y)
input_data_x_old = input_data_x
for i in range(0, input_data_x.shape[1]):
input_data_x = np.delete(input_data_x, i, 1)
diff = high_score - mutinf(100, input_data_x, input_data_y)
bias_num.append(diff)
input_data_x = input_data_x_old
bias_num = np.array(bias_num)
bias_num[bias_num < 0] = 0
return bias_num
def JMI_selection(input_data_x, input_data_y):
max_mi = np.argmax(mutual_info_classif(input_data_x, input_data_y))
input_data_y = np.reshape(input_data_y, (-1, 1))
select_index = []
select_index.append(max_mi)
selected_features = input_data_x[:, max_mi].reshape(-1, 1)
diff = []
for x in range(0, num_of_features):
for i in range(0, input_data_x.shape[1]):
key = 0
for j in range(0, len(select_index)):
if i == select_index[j]:
key = 1
if key == 0:
feature = input_data_x[:, i].reshape(-1, 1)
MI = 0
for k in range(0, selected_features.shape[1]):
MI = MI + mutinf(
100, feature,
np.reshape(selected_features[:, k], (-1, 1)))
diff.append(MI)
else:
diff.append(-99.9)
max_new = np.argmax(diff)
select_index.append(max_new)
del diff[:]
selected_features = np.append(selected_features,
input_data_x[:, max_new].reshape(-1, 1),
axis=1)
return select_index
def GSA_selection_UR(input_data_x, input_data_y):
bias = bias_value(input_data_x, input_data_y)
max_mi = np.argmax(mutual_info_classif(input_data_x, input_data_y))
input_data_y = np.reshape(input_data_y, (-1, 1))
data_new = input_data_x[:, max_mi]
all_data = np.reshape(data_new, (-1, 1))
select_index = []
select_index.append(max_mi)
for i in range(0, num_of_features):
score = []
all_data_old = all_data
for j in range(0, input_data_x.shape[1]):
key = 1
for k in range(0, len(select_index)):
if j == select_index[k]:
key = 0
if key == 1:
all_data = np.append(all_data,
np.reshape(input_data_x[:, j], (-1, 1)),
axis=1)
score.append(
mutinf(100, np.reshape(all_data,
(-1, i + 2)), input_data_y) *
(1 - beta) + beta * bias[j])
else:
score.append(-99.9)
all_data = all_data_old
new_index = np.argmax(score)
select_index.append(new_index)
all_data = np.append(all_data,
np.reshape(input_data_x[:, new_index], (-1, 1)),
axis=1)
return select_index
def JMIM_selection_UR(input_data_x, input_data_y):
bias = bias_value(input_data_x, input_data_y)
max_mi = np.argmax(mutual_info_classif(input_data_x, input_data_y))
input_data_y = np.reshape(input_data_y, (-1, 1))
select_index = []
select_index.append(max_mi)
selected_features = input_data_x[:, max_mi].reshape(-1, 1)
for x in range(0, num_of_features):
all_min = []
for i in range(0, input_data_x.shape[1]):
key = 0
for j in range(0, len(select_index)):
if i == select_index[j]:
key = 1
if key == 0:
feature = input_data_x[:, i]
MI = []
for k in range(0, selected_features.shape[1]):
MI.append(
mutinf(100, np.reshape(feature, (-1, 1)),
np.reshape(selected_features[:, k], (-1, 1))))
all_min.append((1 - beta) * np.min(MI) + beta * bias[i])
else:
all_min.append(-99.9)
max_new = np.argmax(all_min)
select_index.append(max_new)
selected_features = np.append(selected_features,
input_data_x[:, max_new].reshape(-1, 1),
axis=1)
return select_index
def MI_selection(all_data): all_MI = [] original_number = np.unique(all_data.values[:,0]) for i in range(0,len(original_number)): MI = 0 sample_num = original_number[i] for j in range(0,len(original_number)): sample_num_2 = original_number[j] if sample_num_2 != sample_num: P1 = all_data[all_data['Subject No']==sample_num].T P2 = all_data[all_data['Subject No']==sample_num_2].T MI = MI + mutinf(3,P1,P2) all_MI.append(MI) order = np.argsort(all_MI)[::-1] new_order = original_number[order] return new_order
def normalize_trace(tr):
ts, rs = zip(*tr)
ts = np.array(ts) - ts[0]
interp = interp1d(x=ts, y=rs, kind='linear')
return interp(norm_trace_times)
print 'Normalizing reward traces...'
norm_reward_traces = [normalize_trace(tr) for tr in reward_traces]
norm_reward_traces = np.array(norm_reward_traces)
# TODO Consider context as well
# 2. Compute MI between pre/actions, reward values
pre_act_mi = [
mde.mutinf(n_bins=11,
x=pre_actions_contexts,
y=norm_reward_traces[:, i]) for i in range(num_trace_times)
]
post_act_mi = [
mde.mutinf(n_bins=11, x=actions_contexts, y=norm_reward_traces[:, i])
for i in range(num_trace_times)
]
plt.ion()
plt.figure()
plt.title('Instantaneous reward MI vs. time')
plt.plot(norm_trace_times, pre_act_mi, 'r.-', label='prev_act')
plt.plot(norm_trace_times, post_act_mi, 'b.-', label='curr_act')
plt.xlabel('Time after action, t (s)')
plt.ylabel('Mutual Information')
plt.legend(loc='best')