-
Notifications
You must be signed in to change notification settings - Fork 3
/
simulator_save.py
605 lines (537 loc) · 25.6 KB
/
simulator_save.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
import numpy as np
from scipy.special import expit
from sklearn.neighbors import KernelDensity
from sklearn.model_selection import GridSearchCV
from multiprocessing import Pool
class Simulator:
def save_init_state(self):
init_state = np.random.normal(size=self.dim_state, scale=0.1).reshape(-1)
return init_state
def save_confounder_model(self):
confounder = 2 * np.random.binomial(1, p=0.5, size=1) - 1
return confounder
def save_sc2action_model(self, state, confounder):
pa = 0.1*np.sum(state) + 0.9*confounder
pa = expit(pa)
return pa
def save_sa2mediator_model(self, state, action):
pm = 0.1*np.sum(state) + 0.9*(action - 0.5)
pm = expit(pm)
return pm
def save_smc2reward_model(self, state, mediator, confounder, random):
if self.dim_state == 1:
rmean = 0.5*np.clip(confounder, a_min=0, a_max=1)*state[0] + 0.5*np.clip(confounder, a_min=0, a_max=1)*mediator - 0.1*state[0] # new best
elif self.dim_state >= 2:
rmean = 0.5*np.clip(confounder, a_min=0, a_max=1)*np.sum(state) + 0.5*np.clip(confounder, a_min=0, a_max=1)*mediator - 0.1*np.sum(state) # new best
pass
if random:
reward = np.random.normal(size=1, loc=rmean, scale=0.1)
else:
reward = rmean
return reward
def save_smc2nextstate_model(self, state, mediator, confounder):
next_state = np.copy(state)
if self.dim_state >= 2:
next_state = 0.5*np.clip(confounder, a_min=0, a_max=1) * state + 0.5*np.clip(confounder, a_min=0, a_max=1)*mediator - 0.1*state
elif self.dim_state == 1:
next_state[0] = 0.5*np.clip(confounder, a_min=0, a_max=1)*state[0] + 0.5*np.clip(confounder, a_min=0, a_max=1)*mediator - 0.1*state[0]
else:
pass
cov_matrix = 0.1*np.eye(self.dim_state)
next_state = np.random.multivariate_normal(size=1, mean=next_state, cov=cov_matrix)
next_state = next_state.flatten()
return next_state
def standard_init_state(self):
init_state = np.random.normal(size=self.dim_state).reshape(-1)
return init_state
def standard_sc2action_model(self, state, confounder):
pa = 0.1*np.sum(state)
pa = expit(pa)
return pa
def standard_sa2mediator_model(self, state, action):
pm = action
return pm
def standard_smc2reward_model(self, state, mediator, confounder, random):
if self.dim_state == 1:
rmean = state[0] + mediator
pass
if random:
reward = np.random.normal(size=1, loc=rmean, scale=0.25)
else:
reward = rmean
return reward
def standard_smc2nextstate_model(self, state, mediator, confounder):
next_state = np.copy(state)
next_state[0] = mediator - 0.5 + next_state[0]
cov_matrix = 0.25*np.eye(self.dim_state)
next_state = np.random.multivariate_normal(size=1, mean=next_state, cov=cov_matrix)
next_state = next_state.flatten()
return next_state
def toy_init_state(self):
init_state = np.random.binomial(n=1, p=0.5, size=1).reshape(-1)
return init_state
def toy_confounder_model(self):
confounder = 2 * np.random.binomial(n=1, p=0.5, size=1) - 1
return confounder
def toy_sc2action_model(self, state, confounder):
pa = 0.1*np.sum(state) + 0.9*confounder
pa = expit(pa)
return pa
def toy_sa2mediator_model(self, state, action):
pm = 0.1*np.sum(state) + 0.9*action - 0.45
pm = expit(pm)
return pm
def toy_smc2reward_model(self, state, mediator, confounder, random):
rmean = 0.5 * np.clip(confounder, 0, 1) * (state[0] + mediator) - 0.1 * state[0]
rmean = expit(rmean)
if random:
reward = np.random.binomial(n=1, p=rmean, size=1) * 10
else:
reward = rmean
return reward
def toy_smc2nextstate_model(self, state, mediator, confounder):
next_state = 0.5 * np.clip(confounder, 0, 1) * (state[0] + mediator) - 0.1 * state[0]
next_state = expit(next_state)
next_state = np.random.binomial(n=1, p=next_state, size=1)
return next_state
def __init__(self, model_type='save', dim_state=3):
self.dim_state = dim_state
if model_type == 'save':
self.init_state_model = self.save_init_state
self.confounder_model = self.save_confounder_model
self.sc2action_model = self.save_sc2action_model
self.sa2mediator_model = self.save_sa2mediator_model
self.smc2reward_model = self.save_smc2reward_model
self.smc2nextstate_model = self.save_smc2nextstate_model
elif model_type == "standard":
self.init_state_model = self.standard_init_state
self.confounder_model = lambda : 0
self.sc2action_model = self.standard_sc2action_model
self.sa2mediator_model = self.standard_sa2mediator_model
self.smc2reward_model = self.standard_smc2reward_model
self.smc2nextstate_model = self.standard_smc2nextstate_model
elif model_type == "toy":
self.init_state_model = self.toy_init_state
self.confounder_model = self.toy_confounder_model
self.sc2action_model = self.toy_sc2action_model
self.sa2mediator_model = self.toy_sa2mediator_model
self.smc2reward_model = self.toy_smc2reward_model
self.smc2nextstate_model = self.toy_smc2nextstate_model
else:
pass
self.trajectory_list = []
self.target_policy_trajectory_list = []
self.target_policy_state_density_list = None
self.stationary_behaviour_policy_state_density = None
pass
def sample_init_state(self):
init_state = self.init_state_model()
return init_state
def sample_confounder(self):
confounder = self.confounder_model()
return confounder
def logistic_sampler(self, prob):
if prob.ndim == 1:
prob = prob[0]
elif prob.ndim == 2:
prob = prob[0][0]
prob_arr = np.array([1-prob, prob])
random_y = np.random.choice([0, 1], 1, p=prob_arr)
return random_y
def sample_sc2action(self, state, confounder, random=True):
'''
Output: a random action
'''
if random:
random_action = self.logistic_sampler(
self.sc2action_model(state, confounder))
else:
random_action = self.sc2action_model(state, confounder)
return random_action
def sample_sa2mediator(self, state, action):
'''
Output: a random mediator
'''
random_mediator = self.logistic_sampler(
self.sa2mediator_model(state, action))
return random_mediator
def sample_smc2reward(self, state, mediator, confounder, random=True):
random_reward = self.smc2reward_model(
state, mediator, confounder, random=random)
return random_reward
def sample_smc2nextstate(self, state, mediator, confounder):
random_next_state = self.smc2nextstate_model(
state, mediator, confounder)
return random_next_state
def sample_one_trajectory(self, num_time, burn_in):
'''
Output: A list containing 4 elements: state, action, mediator, reward
'''
if burn_in:
burn_in_time = 50
num_time += burn_in_time
init_state = self.sample_init_state()
random_state = np.zeros((num_time+1, self.dim_state))
random_action = np.zeros(num_time)
random_confounder = np.zeros(num_time)
random_mediator = np.zeros(num_time)
random_reward = np.zeros(num_time)
random_state[0] = init_state.reshape(-1)
for i in range(num_time):
random_confounder[i] = self.sample_confounder()
random_action[i] = self.sample_sc2action(
random_state[i], random_confounder[i])
random_mediator[i] = self.sample_sa2mediator(
random_state[i], random_action[i])
random_reward[i] = self.sample_smc2reward(
random_state[i], random_mediator[i], random_confounder[i])
random_state[i+1] = self.sample_smc2nextstate(
random_state[i], random_mediator[i], random_confounder[i])
pass
if burn_in:
valid_index = range(burn_in_time, num_time+1)
random_state = random_state[valid_index]
valid_index = range(burn_in_time, num_time)
random_action = random_action[valid_index]
random_mediator = random_mediator[valid_index]
random_reward = random_reward[valid_index]
random_trajectory = [random_state, random_action, random_mediator, random_reward]
return random_trajectory
def sample_trajectory(self, num_trajectory, num_time, seed, burn_in=False, return_trajectory=False):
tmp_list = self.trajectory_list.copy()
self.trajectory_list = []
np.random.seed(7654321*seed)
for i in range(num_trajectory):
one_trajectory = self.sample_one_trajectory(num_time, burn_in)
self.trajectory_list.append(one_trajectory)
pass
if return_trajectory:
to_return_list = self.trajectory_list.copy()
self.trajectory_list = tmp_list
return to_return_list
def sample_one_target_policy_trajectory(self, num_time, target_policy):
'''
Output: A list containing 4 elements: state, action, mediator, reward
'''
init_state = self.sample_init_state()
random_state = np.zeros((num_time+1, self.dim_state))
random_action = np.zeros(num_time)
random_confounder = np.zeros(num_time)
random_mediator = np.zeros(num_time)
random_reward = np.zeros(num_time)
random_state[0] = init_state.reshape(-1)
for i in range(num_time):
random_confounder[i] = self.sample_confounder()
random_action[i] = target_policy(random_state[i])
# random_mediator[i] = self.sample_action2mediator(random_action[i])
random_mediator[i] = self.sample_sa2mediator(
random_state[i], random_action[i])
random_reward[i] = self.sample_smc2reward(
random_state[i], random_mediator[i], random_confounder[i])
random_state[i+1] = self.sample_smc2nextstate(
random_state[i], random_mediator[i], random_confounder[i])
pass
random_trajectory = [random_state, random_action,
random_mediator, random_reward]
return random_trajectory
def sample_target_policy_trajectory(self, num_trajectory, num_time, target_policy, seed, return_trajectory=False):
tmp_list = self.target_policy_trajectory_list.copy()
self.target_policy_trajectory_list = []
np.random.seed(seed)
for i in range(num_trajectory):
one_trajectory = self.sample_one_target_policy_trajectory(
num_time, target_policy)
self.target_policy_trajectory_list.append(one_trajectory)
pass
if return_trajectory:
to_return_list = self.target_policy_trajectory_list.copy()
self.target_policy_trajectory_list = tmp_list
return to_return_list
def onetrajectory2iid(self, trajectory):
num_time = trajectory[1].shape[0]
s0 = trajectory[0][0]
state = trajectory[0][range(num_time)]
next_state = trajectory[0][range(1, num_time+1)]
trajectory[0] = state
trajectory.append(next_state)
return s0, trajectory
def trajectory2iid(self, trajectory=None):
iid_dataset = []
if trajectory is None:
trajectory_list = self.trajectory_list.copy()
else:
trajectory_list = trajectory.copy()
pass
num_trajectory = len(trajectory_list)
for i in range(num_trajectory):
s0_data, iid_data = self.onetrajectory2iid(trajectory_list[i])
if i == 0:
iid_dataset = iid_data
s0_dataset = s0_data
else:
s0_dataset = np.vstack([s0_dataset, s0_data])
iid_dataset[0] = np.vstack([iid_dataset[0], iid_data[0]])
iid_dataset[4] = np.vstack([iid_dataset[4], iid_data[4]])
iid_dataset[1] = np.append(iid_dataset[1], iid_data[1])
iid_dataset[2] = np.append(iid_dataset[2], iid_data[2])
iid_dataset[3] = np.append(iid_dataset[3], iid_data[3])
pass
pass
self.iid_dataset = {'s0': s0_dataset, 'state': iid_dataset[0],
'action': iid_dataset[1], 'mediator': iid_dataset[2],
'reward': iid_dataset[3], 'next_state': iid_dataset[4]}
if trajectory is not None:
return {'s0': s0_dataset, 'state': iid_dataset[0],
'action': iid_dataset[1], 'mediator': iid_dataset[2],
'reward': iid_dataset[3], 'next_state': iid_dataset[4]}
def estimate_ope(self, target_policy, gamma, max_time=43, mc_s0_time=25, mc_mediator_time=20, burn_in=False, seed=1):
v_value_array = np.zeros(mc_s0_time)
if burn_in:
burn_in_num = 50
max_time += burn_in_num
for j in range(mc_s0_time):
np.random.seed(j+seed)
current_state = self.sample_init_state()
current_state = current_state.reshape(-1)
v_value = 0.0
for i in range(max_time):
action = target_policy(current_state)
confounder = self.sample_confounder()
mediator = self.sample_sa2mediator(current_state, action)
reward = self.sample_smc2reward(
current_state, mediator, confounder, False)
if burn_in:
if i >= burn_in_num:
v_value += np.power(gamma, i-burn_in_num) * reward
else:
v_value += np.power(gamma, i) * reward
current_state = self.sample_smc2nextstate(
current_state, mediator, confounder)
pass
v_value_array[j] = v_value
pass
true_v_value = np.mean(v_value_array)
return true_v_value
def mc_ope(self, j, max_time, gamma, target_policy, seed, burn_in):
np.random.seed(j+seed)
current_state = self.sample_init_state()
current_state = current_state.reshape(-1)
v_value = 0.0
for i in range(max_time):
action = target_policy(current_state)
confounder = self.sample_confounder()
mediator = self.sample_sa2mediator(current_state, action)
reward = self.sample_smc2reward(
current_state, mediator, confounder, False)
if burn_in:
if i >= burn_in_num:
v_value += np.power(gamma, i-burn_in_num) * reward
else:
v_value += np.power(gamma, i) * reward
current_state = self.sample_smc2nextstate(
current_state, mediator, confounder)
pass
return v_value
def estimate_ope_parallel(self, target_policy, gamma, max_time=43, mc_s0_time=25, mc_mediator_time=20, burn_in=False, seed=1, num_process=5):
if burn_in:
burn_in_num = 50
max_time += burn_in_num
seed_list = np.arange(mc_s0_time, dtype='int64').tolist()
seed_offset = (np.ones(mc_s0_time, dtype='int64') * seed).tolist()
max_time_list = (np.ones(mc_s0_time, dtype='int64') * max_time).tolist()
gamma_list = (np.ones(mc_s0_time) * gamma).tolist()
target_policy_list = [target_policy for i in range(mc_s0_time)]
burn_in_list = [burn_in for i in range(mc_s0_time)]
param_list = zip(seed_list, max_time_list, gamma_list,
target_policy_list, seed_offset, burn_in_list)
with Pool(num_process) as p:
v_value_array = p.starmap(self.mc_ope, param_list)
true_v_value = np.mean(v_value_array)
return true_v_value
def estimate_mediated_ope(self, target_policy, gamma, max_time=43, mc_s0_time=25, mc_mediator_time=20):
v_value_array = np.zeros(mc_s0_time)
for j in range(mc_s0_time):
np.random.seed(j)
current_state = self.sample_init_state()
current_state = current_state.reshape(-1)
v_value = 0.0
for i in range(max_time):
confounder = self.sample_confounder()
action = self.sample_sc2action(current_state, confounder)
mediator = self.sample_sa2mediator(
current_state, target_policy(current_state))
reward = self.sample_smc2reward(
current_state, mediator, confounder, False)
v_value += np.power(gamma, i) * reward
current_state = self.sample_smc2nextstate(
current_state, mediator, confounder)
pass
v_value_array[j] = v_value
pass
true_v_value = np.mean(v_value_array)
return true_v_value
def estimate_vfunction_s0(self, s0, target_policy, gamma, max_time=43, mc_mediator_time=20):
v_value_array = np.zeros(mc_mediator_time)
for j in range(mc_mediator_time):
np.random.seed(j)
current_state = np.copy(s0)
current_state = current_state.reshape(-1)
v_value = 0.0
for i in range(max_time):
action = target_policy(current_state)
confounder = self.sample_confounder()
mediator = self.sample_sa2mediator(current_state, action)
reward = self.sample_smc2reward(
current_state, mediator, confounder, random=False)
v_value += np.power(gamma, i) * reward
current_state = self.sample_smc2nextstate(
current_state, mediator, confounder)
pass
v_value_array[j] = v_value
pass
true_v_value = np.mean(v_value_array)
return true_v_value
def estimate_qfunction(self, reward, next_state, target_policy, gamma, max_time=43, mc_mediator_time=20):
v_est = self.estimate_vfunction_s0(
next_state, target_policy, gamma, max_time, mc_mediator_time)
reward_est = reward
q_function_est = v_est * gamma + reward_est
return q_function_est
def estimate_ope_via_qfunction(self, target_policy, gamma, max_time=43, mc_s0_time=25, mc_cp_action_time=200, mc_cp_mediator_time=200):
ope_value_array = np.zeros(mc_s0_time)
action_value = np.array([0.0, 1.0])
action_prob_value = np.array([0.0, 0.0])
mediator_value = np.array([0.0, 1.0])
mediator_prob_value = np.array([0.0, 0.0])
for j in range(mc_s0_time):
np.random.seed(j)
current_state = self.sample_init_state()
current_state = current_state.reshape(-1)
## estimate p(a|s)
mc_action_array = np.zeros(mc_cp_action_time)
for r in range(mc_cp_action_time):
np.random.seed(r)
# confounder = self.sample_confounder()
# mc_action_array[r] = self.sample_sc2action(
# current_state, confounder)
mc_action_array[r] = target_policy(current_state)
pass
for r, one_action in enumerate(action_value):
action_prob_value[r] = np.where(
mc_action_array == one_action)[0].shape[0]
pass
action_prob_value /= float(mc_cp_action_time)
# print(action_prob_value)
## estimate p(m|s, \pi(s))
mc_mediator_array = np.zeros(mc_cp_mediator_time)
for r in range(mc_cp_mediator_time):
np.random.seed(r)
mc_mediator_array[r] = self.sample_sa2mediator(
current_state, target_policy(current_state))
pass
for r, one_mediator in enumerate(mediator_value):
mediator_prob_value[r] = np.where(
mc_mediator_array == one_mediator)[0].shape[0]
pass
mediator_prob_value /= float(mc_cp_mediator_time)
# print(mediator_prob_value)
## compute q value
confounder = self.sample_confounder()
q_value_list = []
for s, one_action in enumerate(action_value):
for t, one_mediator in enumerate(mediator_value):
reward = self.sample_smc2reward(
current_state, one_mediator, confounder, random=False)
next_state = self.sample_smc2nextstate(
current_state, one_mediator, confounder)
q_value = self.estimate_qfunction(
reward, next_state, target_policy, gamma=gamma, max_time=max_time)
q_value *= mediator_prob_value[s]
q_value *= action_prob_value[t]
q_value_list.append(q_value)
pass
pass
q_value_list = np.array(q_value_list)
ope_value_array[j] = np.sum(q_value_list)
pass
true_ope = np.mean(ope_value_array)
return true_ope
def extract_state(self, trajectory_list, num_trajectory):
state_list = []
for i in range(num_trajectory):
state_list.append(trajectory_list[i][0])
pass
state_all = np.vstack(state_list)
return state_all
def estimate_behaviour_policy_state_density(self, num_trajectory, num_time, seed):
num_trajectory = len(self.trajectory_list)
if num_trajectory == 0:
self.sample_trajectory(num_trajectory, num_time, seed)
pass
state_all = self.extract_state(self.trajectory_list, num_trajectory)
params = {'bandwidth': np.logspace(-1, 1, 20)}
grid = GridSearchCV(KernelDensity(), params)
grid.fit(state_all)
print("best bandwidth: {0}".format(grid.best_estimator_.bandwidth))
self.stationary_behaviour_policy_state_density = grid.best_estimator_
pass
def estimate_target_policy_state_density(self, num_trajectory, num_time, target_policy, seed):
num_trajectory = len(self.target_policy_trajectory_list)
if num_trajectory == 0:
self.sample_target_policy_trajectory(
num_trajectory, num_time, target_policy, seed)
pass
state_list = []
for t in range(num_time):
state_list_time_t = []
for i in range(num_trajectory):
state_list_time_t.append(self.target_policy_trajectory_list[i][0][t, :])
pass
print(state_list_time_t)
t_state = np.vstack(state_list_time_t)
state_list.append(t_state)
params = {'bandwidth': np.logspace(-1, 1, 20)}
for t in range(num_time):
grid = GridSearchCV(KernelDensity(), params)
grid.fit(state_list[t])
print("best bandwidth: {0}".format(grid.best_estimator_.bandwidth))
self.target_policy_state_density_list[t] = grid.best_estimator_
pass
pass
def estimate_ratio(self, state, gamma, num_trajectory, num_time, target_policy, max_time=43, seed=1, replace=False):
if self.target_policy_state_density_list is None or replace:
self.estimate_target_policy_state_density(num_trajectory, num_time, target_policy, seed)
if self.stationary_behaviour_policy_state_density is None or replace:
self.estimate_behaviour_policy_state_density(num_trajectory, num_time, seed)
eval_num = state.shape[0]
numerator = np.zeros(eval_num)
for t in range(max_time):
numerator += np.power(gamma, t) * self.target_policy_state_density_list[t].score_samples(state)
pass
denominator = self.stationary_behaviour_policy_state_density.score_samples(state)
ratio_value = numerator / denominator
return ratio_value
def estimate_discrete_ratio(self, num_trajectory, num_time, target_policy, seed, burn_in=False):
trajectory_discrete_ratio = self.sample_trajectory(
num_trajectory, num_time, seed, burn_in=False, return_trajectory=True)
all_state0 = [_trajectory[0] for _trajectory in trajectory_discrete_ratio]
all_state0 = np.hstack(all_state0)
self.obs_prob_array = np.apply_along_axis(np.mean, 1, all_state0)
trajectory_discrete_ratio = self.sample_target_policy_trajectory(
num_trajectory, num_time, target_policy, seed, return_trajectory=True)
all_state = [_trajectory[0] for _trajectory in trajectory_discrete_ratio]
all_state = np.hstack(all_state)
all_state = all_state[:, -1]
self.target_policy_stationary_prob = np.mean(all_state)
def predict_discrete_ratio(self, state, gamma):
denominator = self.target_policy_stationary_prob * state
denominator += (1.0 - self.target_policy_stationary_prob) * (1.0 - state)
numerator = np.array([0.0])
for t, prob in enumerate(self.obs_prob_array):
numerator_part = prob * state
numerator_part += (1.0 - prob) * (1.0 - state)
numerator_part *= np.power(gamma, t)
numerator += numerator_part
pass
ratio = (1.0 - gamma) * numerator / denominator
return ratio