-
Notifications
You must be signed in to change notification settings - Fork 99
/
evaluate_agent.py
921 lines (778 loc) · 31.3 KB
/
evaluate_agent.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
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
import multiprocessing
import time
from copy import deepcopy
from multiprocessing import Pool
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from scipy.stats import beta
import recogym
from recogym import (
AgentInit,
AgentStats,
Configuration,
EvolutionCase,
RoiMetrics,
TrainingApproach
)
from recogym.agents import EpsilonGreedy, epsilon_greedy_args
from .envs.context import DefaultContext
from .envs.observation import Observation
from .envs.session import OrganicSessions
EpsilonDelta = .02
EpsilonSteps = 6 # Including epsilon = 0.0.
EpsilonPrecision = 2
EvolutionEpsilons = (0.00, 0.01, 0.02, 0.03, 0.05, 0.08)
GraphCTRMin = 0.009
GraphCTRMax = 0.021
# from Keras
def to_categorical(y, num_classes=None, dtype='float32'):
y = np.array(y, dtype='int')
input_shape = y.shape
if input_shape and input_shape[-1] == 1 and len(input_shape) > 1:
input_shape = tuple(input_shape[:-1])
y = y.ravel()
if not num_classes:
num_classes = np.max(y) + 1
n = y.shape[0]
categorical = np.zeros((n, num_classes), dtype=dtype)
categorical[np.arange(n), y] = 1
output_shape = input_shape + (num_classes,)
categorical = np.reshape(categorical, output_shape)
return categorical
def evaluate_agent(
env,
agent,
num_initial_train_users=100,
num_step_users=1000,
num_steps=10,
training_approach=TrainingApproach.ALL_DATA,
sliding_window_samples=10000):
initial_agent = deepcopy(agent)
unique_user_id = 0
for u in range(num_initial_train_users):
env.reset(unique_user_id + u)
agent.reset()
new_observation, reward, done, _ = env.step(None)
while True:
old_observation = new_observation
action, new_observation, reward, done, _ = env.step_offline(
new_observation, reward, False
)
agent.train(old_observation, action, reward, done)
if done:
break
unique_user_id += num_initial_train_users
rewards = {
EvolutionCase.SUCCESS: [],
EvolutionCase.SUCCESS_GREEDY: [],
EvolutionCase.FAILURE: [],
EvolutionCase.FAILURE_GREEDY: [],
EvolutionCase.ACTIONS: dict()
}
training_agent = deepcopy(agent)
samples = 0
for action_id in range(env.config.num_products):
rewards[EvolutionCase.ACTIONS][action_id] = [0]
for step in range(num_steps):
successes = 0
successes_greedy = 0
failures = 0
failures_greedy = 0
for u in range(num_step_users):
env.reset(unique_user_id + u)
agent.reset()
new_observation, reward, done, _ = env.step(None)
while not done:
old_observation = new_observation
action = agent.act(old_observation, reward, done)
new_observation, reward, done, info = env.step(action['a'])
samples += 1
should_update_training_data = False
if training_approach == TrainingApproach.ALL_DATA or training_approach == TrainingApproach.LAST_STEP:
should_update_training_data = True
elif training_approach == TrainingApproach.SLIDING_WINDOW_ALL_DATA:
should_update_training_data = samples % sliding_window_samples == 0
elif training_approach == TrainingApproach.ALL_EXPLORATION_DATA:
should_update_training_data = not action['greedy']
elif training_approach == TrainingApproach.SLIDING_WINDOW_EXPLORATION_DATA:
should_update_training_data = (not action[
'greedy']) and samples % sliding_window_samples == 0
else:
assert False, f"Unknown Training Approach: {training_approach}"
if should_update_training_data:
training_agent.train(old_observation, action, reward, done)
if reward:
successes += 1
if 'greedy' in action and action['greedy']:
successes_greedy += 1
rewards[EvolutionCase.ACTIONS][action['a']][-1] += 1
else:
if 'greedy' in action and action['greedy']:
failures_greedy += 1
failures += 1
unique_user_id += num_step_users
agent = training_agent
for action_id in range(env.config.num_products):
rewards[EvolutionCase.ACTIONS][action_id].append(0)
if training_approach == TrainingApproach.LAST_STEP:
training_agent = deepcopy(initial_agent)
else:
training_agent = deepcopy(agent)
rewards[EvolutionCase.SUCCESS].append(successes)
rewards[EvolutionCase.SUCCESS_GREEDY].append(successes_greedy)
rewards[EvolutionCase.FAILURE].append(failures)
rewards[EvolutionCase.FAILURE_GREEDY].append(failures_greedy)
return rewards
def build_agent_init(agent_key, ctor, def_args):
return {
agent_key: {
AgentInit.CTOR: ctor,
AgentInit.DEF_ARGS: def_args,
}
}
def _collect_stats(args):
"""
Function that is executed in a separate process.
:param args: arguments of the process to be executed.
:return: a vector of CTR for these confidence values:
0th: Q0.500
1st: Q0.025
snd: Q0.975
"""
start = time.time()
print(f"START: Num of Users: {args['num_offline_users']}")
stats = recogym.test_agent(
deepcopy(args['env']),
deepcopy(args['agent']),
args['num_offline_users'],
args['num_online_users'],
args['num_organic_offline_users'],
args['num_epochs'],
args['epoch_with_random_reset'],
args['with_cache'],
)
print(f"END: Num of Offline Users: {args['num_offline_users']} ({time.time() - start}s)")
return stats
def gather_agent_stats(
env,
env_args,
extra_env_args,
agents_init_data,
user_samples=(100, 1000, 2000, 3000, 5000, 8000, 10000, 13000, 14000, 15000),
num_online_users: int = 15000,
num_epochs: int = 1,
epoch_with_random_reset: bool = False,
num_organic_offline_users: int = 100,
with_cache: bool = False
):
"""
The function that gathers Agents statistics via evaluating Agent performance
under different Environment conditions.
:param env: the Environment where some changes should be introduced and where Agent stats should
be gathered.
:param env_args: Environment arguments (default ones).
:param extra_env_args: extra Environment conditions those alter default values.
:param agents_init_data: Agent initialisation data.
This is a dictionary that has the following structure:
{
'<Agent Name>': {
AgentInit.CTOR: <Constructor>,
AgentInit.DEF_ARG: <Default Arguments>,
}
}
:param user_samples: Number of Offline Users i.e. Users used to train a Model.
:param num_online_users: Number of Online Users i.e. Users used to validate a Model.
:param num_epochs: how many different epochs should be tried to gather stats?
:param epoch_with_random_reset: should be a Random Seed reset at each new epoch?
:param num_organic_offline_users: how many Organic only users should be used for training.
:param with_cache: is the cache used for training data or not?
:return: a dictionary with stats
{
AgentStats.SAMPLES: [<vector of training offline users used to train a model>]
AgentStats.AGENTS: {
'<Agent Name>': {
AgentStats.Q0_025: [],
AgentStats.Q0_500: [],
AgentStats.Q0_975: [],
}
}
}
"""
new_env_args = {
**env_args,
**extra_env_args,
}
new_env = deepcopy(env)
new_env.init_gym(new_env_args)
agents = build_agents(agents_init_data, new_env_args)
agent_stats = {
AgentStats.SAMPLES: user_samples,
AgentStats.AGENTS: dict(),
}
for agent_key in agents:
print(f"Agent: {agent_key}")
stats = {
AgentStats.Q0_025: [],
AgentStats.Q0_500: [],
AgentStats.Q0_975: [],
}
with Pool(processes=multiprocessing.cpu_count()) as pool:
argss = [
{
'env': new_env,
'agent': agents[agent_key],
'num_offline_users': num_offline_users,
'num_online_users': num_online_users,
'num_organic_offline_users': num_organic_offline_users,
'num_epochs': num_epochs,
'epoch_with_random_reset': epoch_with_random_reset,
'with_cache': with_cache
}
for num_offline_users in user_samples
]
for result in (
[_collect_stats(args) for args in argss]
if num_epochs == 1 else
pool.map(_collect_stats, argss)
):
stats[AgentStats.Q0_025].append(result[1])
stats[AgentStats.Q0_500].append(result[0])
stats[AgentStats.Q0_975].append(result[2])
agent_stats[AgentStats.AGENTS][agent_key] = stats
return agent_stats
def build_agents(agents_init_data, new_env_args):
agents = dict()
for agent_key in agents_init_data:
agent_init_data = agents_init_data[agent_key]
ctor = agent_init_data[AgentInit.CTOR]
def_args = agent_init_data[AgentInit.DEF_ARGS]
agents[agent_key] = ctor(
Configuration({
**def_args,
**new_env_args,
})
)
return agents
def generate_epsilons(epsilon_step=EpsilonDelta, iterations=EpsilonSteps):
return [0.00, 0.01, 0.02, 0.03, 0.05, 0.08]
def format_epsilon(epsilon):
return ("{0:." + f"{EpsilonPrecision}" + "f}").format(round(epsilon, EpsilonPrecision))
def _collect_evolution_stats(args):
"""
Function that is executed in a separate process.
:param args: arguments of the process to be executed.
:return: a dictionary of Success/Failures of applying an Agent.
"""
start = time.time()
epsilon = args['epsilon']
epsilon_key = format_epsilon(epsilon)
print(f"START: ε = {epsilon_key}")
num_evolution_steps = args['num_evolution_steps']
rewards = recogym.evaluate_agent(
deepcopy(args['env']),
args['agent'],
args['num_initial_train_users'],
args['num_step_users'],
num_evolution_steps,
args['training_approach']
)
assert (len(rewards[EvolutionCase.SUCCESS]) == len(rewards[EvolutionCase.FAILURE]))
assert (len(rewards[EvolutionCase.SUCCESS]) == num_evolution_steps)
print(f"END: ε = {epsilon_key} ({time.time() - start}s)")
return {
epsilon_key: {
EvolutionCase.SUCCESS: rewards[EvolutionCase.SUCCESS],
EvolutionCase.SUCCESS_GREEDY: rewards[EvolutionCase.SUCCESS_GREEDY],
EvolutionCase.FAILURE: rewards[EvolutionCase.FAILURE],
EvolutionCase.FAILURE_GREEDY: rewards[EvolutionCase.FAILURE_GREEDY],
EvolutionCase.ACTIONS: rewards[EvolutionCase.ACTIONS]
}
}
def gather_exploration_stats(
env,
env_args,
extra_env_args,
agents_init_data,
training_approach,
num_initial_train_users=1000,
num_step_users=1000,
epsilons=EvolutionEpsilons,
num_evolution_steps=6
):
"""
A helper function that collects data regarding Agents evolution
under different values of epsilon for Epsilon-Greedy Selection Policy.
:param env: The Environment where evolution should be applied;
every time when a new step of the evolution is applied, the Environment is deeply copied
thus the Environment does not interferes with evolution steps.
:param env_args: Environment arguments (default ones).
:param extra_env_args: extra Environment conditions those alter default values.
:param agents_init_data: Agent initialisation data.
This is a dictionary that has the following structure:
{
'<Agent Name>': {
AgentInit.CTOR: <Constructor>,
AgentInit.DEF_ARG: <Default Arguments>,
}
}
:param training_approach: A training approach applied in verification;
for mode details look at `TrainingApproach' enum.
:param num_initial_train_users: how many users' data should be used
to train an initial model BEFORE evolution steps.
:param num_step_users: how many users' data should be used
at each evolution step.
:param epsilons: a list of epsilon values.
:param num_evolution_steps: how many evolution steps should be applied
for an Agent with Epsilon-Greedy Selection Policy.
:return a dictionary of Agent evolution statistics in the form:
{
'Agent Name': {
'Epsilon Values': {
EvolutionCase.SUCCESS: [an array of clicks (for each ith step of evolution)]
EvolutionCase.FAILURE: [an array of failure to draw a click (for each ith step of evolution)]
}
}
}
"""
# A dictionary that stores all data of Agent evolution statistics.
# Key is Agent Name, value is statistics.
agent_evolution_stats = dict()
new_env_args = {
**env_args,
**extra_env_args,
}
new_env = deepcopy(env)
new_env.init_gym(new_env_args)
agents = build_agents(agents_init_data, new_env_args)
for agent_key in agents:
print(f"Agent: {agent_key}")
agent_stats = dict()
with Pool(processes=multiprocessing.cpu_count()) as pool:
for result in pool.map(
_collect_evolution_stats,
[
{
'epsilon': epsilon,
'env': new_env,
'agent': EpsilonGreedy(
Configuration({
**epsilon_greedy_args,
**new_env_args,
'epsilon': epsilon,
}),
deepcopy(agents[agent_key])
),
'num_initial_train_users': num_initial_train_users,
'num_step_users': num_step_users,
'num_evolution_steps': num_evolution_steps,
'training_approach': training_approach,
}
for epsilon in epsilons
]
):
agent_stats = {
**agent_stats,
**result,
}
agent_evolution_stats[agent_key] = agent_stats
return agent_evolution_stats
def plot_agent_stats(agent_stats):
_, ax = plt.subplots(
1,
1,
figsize=(16, 8)
)
user_samples = agent_stats[AgentStats.SAMPLES]
for agent_key in agent_stats[AgentStats.AGENTS]:
stats = agent_stats[AgentStats.AGENTS][agent_key]
ax.fill_between(
user_samples,
stats[AgentStats.Q0_975],
stats[AgentStats.Q0_025],
alpha=.05
)
ax.plot(user_samples, stats[AgentStats.Q0_500])
ax.set_xlabel('Samples #')
ax.set_ylabel('CTR')
ax.legend([
"$C^{CTR}_{0.5}$: " + f"{agent_key}" for agent_key in agent_stats[AgentStats.AGENTS]
])
plt.show()
def plot_evolution_stats(
agent_evolution_stats,
max_agents_per_row=2,
epsilons=EvolutionEpsilons,
plot_min=GraphCTRMin,
plot_max=GraphCTRMax
):
figs, axs = plt.subplots(
int(len(agent_evolution_stats) / max_agents_per_row),
max_agents_per_row,
figsize=(16, 10),
squeeze=False
)
labels = [("$\epsilon=$" + format_epsilon(epsilon)) for epsilon in epsilons]
for (ix, agent_key) in enumerate(agent_evolution_stats):
ax = axs[int(ix / max_agents_per_row), int(ix % max_agents_per_row)]
agent_evolution_stat = agent_evolution_stats[agent_key]
ctr_means = []
for epsilon in epsilons:
epsilon_key = format_epsilon(epsilon)
evolution_stat = agent_evolution_stat[epsilon_key]
steps = []
ms = []
q0_025 = []
q0_975 = []
assert (len(evolution_stat[EvolutionCase.SUCCESS]) == len(
evolution_stat[EvolutionCase.FAILURE]))
for step in range(len(evolution_stat[EvolutionCase.SUCCESS])):
steps.append(step)
successes = evolution_stat[EvolutionCase.SUCCESS][step]
failures = evolution_stat[EvolutionCase.FAILURE][step]
ms.append(beta.ppf(0.5, successes + 1, failures + 1))
q0_025.append(beta.ppf(0.025, successes + 1, failures + 1))
q0_975.append(beta.ppf(0.975, successes + 1, failures + 1))
ctr_means.append(np.mean(ms))
ax.fill_between(
range(len(steps)),
q0_975,
q0_025,
alpha=.05
)
ax.plot(steps, ms)
ctr_means_mean = np.mean(ctr_means)
ctr_means_div = np.sqrt(np.var(ctr_means))
ax.set_title(
f"Agent: {agent_key}\n"
+ "$\hat{Q}^{CTR}_{0.5}="
+ "{0:.5f}".format(round(ctr_means_mean, 5))
+ "$, "
+ "$\hat{\sigma}^{CTR}_{0.5}="
+ "{0:.5f}".format(round(ctr_means_div, 5))
+ "$"
)
ax.legend(labels)
ax.set_ylabel('CTR')
ax.set_ylim([plot_min, plot_max])
plt.subplots_adjust(hspace=.5)
plt.show()
def plot_heat_actions(
agent_evolution_stats,
epsilons=EvolutionEpsilons
):
max_epsilons_per_row = len(epsilons)
the_first_agent = next(iter(agent_evolution_stats.values()))
epsilon_steps = len(the_first_agent)
rows = int(len(agent_evolution_stats) * epsilon_steps / max_epsilons_per_row)
figs, axs = plt.subplots(
int(len(agent_evolution_stats) * epsilon_steps / max_epsilons_per_row),
max_epsilons_per_row,
figsize=(16, 4 * rows),
squeeze=False
)
for (ix, agent_key) in enumerate(agent_evolution_stats):
agent_evolution_stat = agent_evolution_stats[agent_key]
for (jx, epsilon_key) in enumerate(agent_evolution_stat):
flat_index = ix * epsilon_steps + jx
ax = axs[int(flat_index / max_epsilons_per_row), int(flat_index % max_epsilons_per_row)]
evolution_stat = agent_evolution_stat[epsilon_key]
action_stats = evolution_stat[EvolutionCase.ACTIONS]
total_actions = len(action_stats)
heat_data = []
for kx in range(total_actions):
heat_data.append(action_stats[kx])
heat_data = np.array(heat_data)
im = ax.imshow(heat_data)
ax.set_yticks(np.arange(total_actions))
ax.set_yticklabels([f"{action_id}" for action_id in range(total_actions)])
ax.set_title(f"Agent: {agent_key}\n$\epsilon=${epsilon_key}")
_ = ax.figure.colorbar(im, ax=ax)
plt.show()
def plot_roi(
agent_evolution_stats,
epsilons=EvolutionEpsilons,
max_agents_per_row=2
):
"""
A helper function that calculates Return of Investment (ROI) for applying Epsilon-Greedy Selection Policy.
:param agent_evolution_stats: statistic about Agent evolution collected in `build_exploration_data'.
:param epsilons: a list of epsilon values.
:param max_agents_per_row: how many graphs should be drawn per a row
:return: a dictionary of Agent ROI after applying Epsilon-Greedy Selection Strategy in the following form:
{
'Agent Name': {
'Epsilon Value': {
Metrics.ROI: [an array of ROIs for each ith step (starting from 1st step)]
}
}
}
"""
figs, axs = plt.subplots(
int(len(agent_evolution_stats) / max_agents_per_row),
max_agents_per_row,
figsize=(16, 8),
squeeze=False
)
labels = [("$\epsilon=$" + format_epsilon(epsilon)) for epsilon in epsilons if epsilon != 0.0]
agent_roi_stats = dict()
for (ix, agent_key) in enumerate(agent_evolution_stats):
ax = axs[int(ix / max_agents_per_row), int(ix % max_agents_per_row)]
agent_stat = agent_evolution_stats[agent_key]
zero_epsilon_key = format_epsilon(0)
zero_epsilon = agent_stat[zero_epsilon_key]
zero_success_evolutions = zero_epsilon[EvolutionCase.SUCCESS]
zero_failure_evolutions = zero_epsilon[EvolutionCase.FAILURE]
assert (len(zero_success_evolutions))
agent_stats = dict()
roi_mean_means = []
for epsilon in generate_epsilons():
if zero_epsilon_key == format_epsilon(epsilon):
continue
epsilon_key = format_epsilon(epsilon)
agent_stats[epsilon_key] = {
RoiMetrics.ROI_0_025: [],
RoiMetrics.ROI_MEAN: [],
RoiMetrics.ROI_0_975: [],
}
epsilon_evolutions = agent_stat[epsilon_key]
success_greedy_evolutions = epsilon_evolutions[EvolutionCase.SUCCESS_GREEDY]
failure_greedy_evolutions = epsilon_evolutions[EvolutionCase.FAILURE_GREEDY]
assert (len(success_greedy_evolutions) == len(failure_greedy_evolutions))
assert (len(zero_success_evolutions) == len(success_greedy_evolutions))
steps = []
roi_means = []
for step in range(1, len(epsilon_evolutions[EvolutionCase.SUCCESS])):
previous_zero_successes = zero_success_evolutions[step - 1]
previous_zero_failures = zero_failure_evolutions[step - 1]
current_zero_successes = zero_success_evolutions[step]
current_zero_failures = zero_failure_evolutions[step]
current_epsilon_greedy_successes = success_greedy_evolutions[step]
current_epsilon_greedy_failures = failure_greedy_evolutions[step]
def roi_with_confidence_interval(
epsilon,
previous_zero_successes,
previous_zero_failures,
current_zero_successes,
current_zero_failures,
current_epsilon_greedy_successes,
current_epsilon_greedy_failures
):
def roi_formulae(
epsilon,
previous_zero,
current_zero,
current_epsilon_greedy
):
current_gain = current_epsilon_greedy / (1 - epsilon) - current_zero
roi = current_gain / (epsilon * previous_zero)
return roi
return {
RoiMetrics.ROI_SUCCESS: roi_formulae(
epsilon,
previous_zero_successes,
current_zero_successes,
current_epsilon_greedy_successes
),
RoiMetrics.ROI_FAILURE: roi_formulae(
epsilon,
previous_zero_failures,
current_zero_failures,
current_epsilon_greedy_failures
)
}
roi_mean = roi_with_confidence_interval(
epsilon,
previous_zero_successes,
previous_zero_failures,
current_zero_successes,
current_zero_failures,
current_epsilon_greedy_successes,
current_epsilon_greedy_failures
)[RoiMetrics.ROI_SUCCESS]
agent_stats[epsilon_key][RoiMetrics.ROI_MEAN].append(roi_mean)
roi_means.append(roi_mean)
steps.append(step)
roi_mean_means.append(np.mean(roi_means))
ax.plot(steps, roi_means)
roi_means_mean = np.mean(roi_mean_means)
roi_means_div = np.sqrt(np.var(roi_mean_means))
ax.set_title(
"$ROI_{t+1}$ of Agent: " + f"'{agent_key}'\n"
+ "$\hat{\mu}_{ROI}="
+ "{0:.5f}".format(round(roi_means_mean, 5))
+ "$, "
+ "$\hat{\sigma}_{ROI}="
+ "{0:.5f}".format(round(roi_means_div, 5))
+ "$"
)
ax.legend(labels, loc=10)
ax.set_ylabel('ROI')
agent_roi_stats[agent_key] = agent_stats
plt.subplots_adjust(hspace=.5)
plt.show()
return agent_roi_stats
def verify_agents(env, number_of_users, agents):
stat = {
'Agent': [],
'0.025': [],
'0.500': [],
'0.975': [],
}
for agent_id in agents:
stat['Agent'].append(agent_id)
data = deepcopy(env).generate_logs(number_of_users, agents[agent_id])
bandits = data[data['z'] == 'bandit']
successes = bandits[bandits['c'] == 1].shape[0]
failures = bandits[bandits['c'] == 0].shape[0]
stat['0.025'].append(beta.ppf(0.025, successes + 1, failures + 1))
stat['0.500'].append(beta.ppf(0.500, successes + 1, failures + 1))
stat['0.975'].append(beta.ppf(0.975, successes + 1, failures + 1))
return pd.DataFrame().from_dict(stat)
def evaluate_IPS(agent, reco_log):
ee = []
for u in range(max(reco_log.u)):
t = np.array(reco_log[reco_log['u'] == u].t)
v = np.array(reco_log[reco_log['u'] == u].v)
a = np.array(reco_log[reco_log['u'] == u].a)
c = np.array(reco_log[reco_log['u'] == u].c)
z = list(reco_log[reco_log['u'] == u].z)
ps = np.array(reco_log[reco_log['u'] == u].ps)
jj = 0
session = OrganicSessions()
agent.reset()
while True:
if jj >= len(z):
break
if z[jj] == 'organic':
session.next(DefaultContext(t[jj], u), int(v[jj]))
else:
prob_policy = agent.act(Observation(DefaultContext(t[jj], u), session), 0, False)[
'ps-a']
if prob_policy!=():
ee.append(c[jj] * prob_policy[int(a[jj])] / ps[jj])
session = OrganicSessions()
jj += 1
return ee
def evaluate_SNIPS(agent, reco_log):
rewards = []
p_ratio = []
for u in range(max(reco_log.u)):
t = np.array(reco_log[reco_log['u'] == u].t)
v = np.array(reco_log[reco_log['u'] == u].v)
a = np.array(reco_log[reco_log['u'] == u].a)
c = np.array(reco_log[reco_log['u'] == u].c)
z = list(reco_log[reco_log['u'] == u].z)
ps = np.array(reco_log[reco_log['u'] == u].ps)
jj = 0
session = OrganicSessions()
agent.reset()
while True:
if jj >= len(z):
break
if z[jj] == 'organic':
session.next(DefaultContext(t[jj], u), int(v[jj]))
else:
prob_policy = agent.act(Observation(DefaultContext(t[jj], u), session), 0, False)[
'ps-a']
rewards.append(c[jj])
p_ratio.append(prob_policy[int(a[jj])] / ps[jj])
session = OrganicSessions()
jj += 1
return rewards, p_ratio
def verify_agents_IPS(reco_log, agents):
stat = {
'Agent': [],
'0.025': [],
'0.500': [],
'0.975': [],
}
for agent_id in agents:
ee = evaluate_IPS(agents[agent_id], reco_log)
mean_ee = np.mean(ee)
se_ee = np.std(ee) / np.sqrt(len(ee))
stat['Agent'].append(agent_id)
stat['0.025'].append(mean_ee - 2 * se_ee)
stat['0.500'].append(mean_ee)
stat['0.975'].append(mean_ee + 2 * se_ee)
return pd.DataFrame().from_dict(stat)
def verify_agents_SNIPS(reco_log, agents):
stat = {
'Agent': [],
'0.025': [],
'0.500': [],
'0.975': [],
}
for agent_id in agents:
rewards, p_ratio = evaluate_SNIPS(agents[agent_id], reco_log)
ee = np.asarray(rewards) * np.asarray(p_ratio)
mean_ee = np.sum(ee) / np.sum(p_ratio)
se_ee = np.std(ee) / np.sqrt(len(ee))
stat['Agent'].append(agent_id)
stat['0.025'].append(mean_ee - 2 * se_ee)
stat['0.500'].append(mean_ee)
stat['0.975'].append(mean_ee + 2 * se_ee)
return pd.DataFrame().from_dict(stat)
def evaluate_recall_at_k(agent, reco_log, k=5):
hits = []
for u in range(max(reco_log.u)):
t = np.array(reco_log[reco_log['u'] == u].t)
v = np.array(reco_log[reco_log['u'] == u].v)
a = np.array(reco_log[reco_log['u'] == u].a)
c = np.array(reco_log[reco_log['u'] == u].c)
z = list(reco_log[reco_log['u'] == u].z)
ps = np.array(reco_log[reco_log['u'] == u].ps)
jj = 0
session = OrganicSessions()
agent.reset()
while True:
if jj >= len(z):
break
if z[jj] == 'organic':
session.next(DefaultContext(t[jj], u), int(v[jj]))
else:
prob_policy = agent.act(Observation(DefaultContext(t[jj], u), session), 0, False)[
'ps-a']
# Does the next session exist?
if (jj + 1) < len(z):
# Is the next session organic?
if z[jj + 1] == 'organic':
# Whas there no click for this bandit event?
if not c[jj]:
# Generate a top-K from the probability distribution over all actions
top_k = set(np.argpartition(prob_policy, -k)[-k:])
# Is the next seen item in the top-K?
if v[jj + 1] in top_k:
hits.append(1)
else:
hits.append(0)
session = OrganicSessions()
jj += 1
return hits
def verify_agents_recall_at_k(reco_log, agents, k=5):
stat = {
'Agent': [],
'0.025': [],
'0.500': [],
'0.975': [],
}
for agent_id in agents:
hits = evaluate_recall_at_k(agents[agent_id], reco_log, k=k)
mean_hits = np.mean(hits)
se_hits = np.std(hits) / np.sqrt(len(hits))
stat['Agent'].append(agent_id)
stat['0.025'].append(mean_hits - 2 * se_hits)
stat['0.500'].append(mean_hits)
stat['0.975'].append(mean_hits + 2 * se_hits)
return pd.DataFrame().from_dict(stat)
def plot_verify_agents(result):
fig, ax = plt.subplots()
ax.set_title('CTR Estimate for Different Agents')
plt.errorbar(result['Agent'],
result['0.500'],
yerr=(result['0.500'] - result['0.025'],
result['0.975'] - result['0.500']),
fmt='o',
capsize=4)
plt.xticks(result['Agent'], result['Agent'], rotation='vertical')
return fig