-
Notifications
You must be signed in to change notification settings - Fork 0
/
algos.py
359 lines (273 loc) · 14.3 KB
/
algos.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
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from tqdm import tqdm
import operator
import argparse
import sys
import matplotlib.pyplot as plt
plt.style.use('data/plots_paper.mplstyle')
import pathlib
class algs:
def __init__(self, theta=67, alpha=3, sigma=2)
assert(theta <= 147 and theta >=0)
# True theta
self.theta = theta
self.alpha = alpha
self.sigma = sigma
# Load data
self.train_data_rating = np.load('preproc/genres/train_data_rating.npy')
#For regret computation
self.test_data_rating = np.load('preproc/genres/test_data_rating.npy')
# Test data
self.test_data_df = pd.read_pickle('genres/test_data_with_id.npy')
# Pulling out relevant columns for quicker lookup
self.test_data = np.concatenate([np.array(test_data.Rating.tolist())[:,None], \
np.array(test_data.Genre_Col.tolist())[:, None], \
np.array(test_data.Meta_User_Col.tolist())[:, None]], \
axis=1)
# Initialize mu [# meta-users, # genres]
mu_ = np.zeros((147, 18))
mu_test = np.zeros((147, 18))
for i in range(len(self.train_data_rating)):
mu_[i,:] = self.train_data_rating[i].values()
mu_test[i,:] = self.test_data_rating[i].values()
theta_set = list(range(147))
# Clear low counts
self.mu, self.mu_test, self.theta_set = self._clear_low_counts(mu_, mu_test, theta_set)
self.numArms = mu.shape[1]
self.bestArm = np.argmax(mu[self.theta, :])
# Re-define theta as updated index - can be removed based on how theta is picked
self.theta = self._index_of(self.theta)
def _index_of(self):
return np.where(self.theta_set == self.theta)[0][0]
def _clear_low_counts(self, mu, mu_test, theta_set):
#These have <250 counts: As selected from the analysis code
remove_metau = [2, 3, 4, 5, 6, 7, 8, 9, 11, 12, 13, 14, 15, 16, 17, 18,\
20, 29, 34, 52, 55, 71, 73, 88, 94, 103, 109, 115, 124, 130, 131, 134, \
135, 136, 137, 144,145]
mu = np.delete(mu, (remove_metau), axis = 0)
zero_counts_remove = np.where(mu == 0)[0]
mu = np.delete(mu, (zero_counts_remove), axis = 0)
mu_test = np.delete(mu_test, (remove_metau), axis = 0)
mu_test = np.delete(mu_test, (zero_counts_remove), axis = 0)
theta_set = np.delete(theta_set, remove_metau)
theta_set = np.delete(theta_set, zero_counts_remove)
return mu, mu_test, theta_set
def generate_sample(self, arm):
genre_idx = self.test_data[:,1] == arm
meta_user_idx = self.test_data[:,2] == self.theta
ratings = self.test_data[:, genre_idx, meta_user_idx][:,0]
if ratings.shape[0] == 0:
return 0
else:
return np.random.choice(ratings)
def confidence_set_intersection(self, empReward, numPulls):
# First 5 steps of the algorithm
T = np.sum(numPulls)
theta_hat = set()
isCompetitive = np.zeros(self.numArms)
#Confidence set construction
with np.errstate(divide='ignore'): # for first few iterations with 0 numpPulls
for theta in self.theta_set:
bound = np.sqrt(2*self.alpha*(self.sigma**2)*np.log(T)/numPulls)
if all(np.abs(self.mu[theta, :] - empReward) <= bound):
theta_hat.add(theta)
if len(theta_hat) == 0:
return np.ones(self.numArms)
# Competitive set
max_mu = np.max(self.mu, axis=1)
for arm in range(self.numArms):
if any(self.mu[:, arm] == max_mu):
isCompetitive[arm] = 1
return isCompetitive, theta_hat
def next_arm_selection(self, rewards):
return np.random.choice(np.flatnonzero(rewards == rewards.max()))
def UCB(self, num_iterations, totalRounds):
avg_regret = np.zero((num_iterations, totalRounds))
for iteration in range(num_iterations):
numPulls = np.zeros(numArms)
empReward = np.zeros(numArms)
regret = np.zeros(totalRounds)
for t in range(totalRounds):
with np.errstate(divide='ignore'):
ucb = self.UCBSample(empReward, numPulls)
next_arm = self.next_arm_selection(ucb)
#Generate reward, update pulls and empirical reward
reward_sample = self.generate_sample(next_arm)
empReward[next_arm] = (empReward[next_arm]*numPulls[next_arm] + reward_sample)/(numPulls[next_arm] + 1)
numPulls[next_arm] = numPulls[next_arm] + 1
#Evaluate regret
regret[t] = self.mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
avg_regret[iteration, :] = regret
return avg_regret
def UCBSample(self, empReward, numPulls):
return empReward + np.sqrt(2*self.alpha*(self.sigma**2)*np.log(np.sum(numPulls))/numPulls)
def ThompsonSample(self, empiricalMean, numPulls):
sampleArm = np.random.normal(empiricalMean, np.sqrt(self.sigma**2/numPulls))
return sampleArm
def TS(self, num_iterations, totalRounds):
avg_regret = np.zero((num_iterations, totalRounds))
for iteration in range(num_iterations):
numPulls = np.zeros(numArms)
empReward = np.zeros(numArms)
regret = np.zeros(totalRounds)
for t in range(totalRounds):
#Initialise by pulling each arm once
if t < numArms:
numPulls[t] += 1
assert numPulls[t] == 1
reward = self.generate_sample(t)
empReward[t] = reward
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
continue
thompson = ThompsonSample(empReward,numPulls)
next_arm = self.next_arm_selection(thompson)
#Generate reward, update pulls and empirical reward
reward_sample = self.generate_sample( next_arm )
empReward[next_arm] = (empReward[next_arm]*numPulls[next_arm] + reward_sample)/(numPulls[next_arm] + 1)
numPulls[next_arm] = numPulls[next_arm] + 1
#Evaluate regret
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
avg_regret[iteration, :] = regret
return avg_regret
def UCB_C(self, num_iterations, totalRounds):
avg_regret = np.zero((num_iterations, totalRounds))
for iteration in range(num_iterations):
numPulls = np.zeros(numArms)
empReward = np.zeros(numArms)
regret = np.zeros(totalRounds)
for t in range(totalRounds):
isCompetitive, _ = self.confidence_set_intersection(empReward, numPulls)
with np.errstate(divide='ignore'):
ucb = self.UCBSample(empReward, numPulls)
if isCompetitive.sum() == 0:
next_arm = =np.random.randint(0,self.numArms)
else:
self.next_arm_selection(ucb*isCompetitive)
#Generate reward, update pulls and empirical reward
reward_sample = self.generate_sample(next_arm)
empReward[next_arm] = (empReward[next_arm]*numPulls[next_arm] + reward_sample)/(numPulls[next_arm] + 1)
numPulls[next_arm] = numPulls[next_arm] + 1
#Evaluate regret
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
avg_regret[iteration, :] = regret
return avg_regret
def TS_C(self, num_iterations, totalRounds):
avg_regret = np.zero((num_iterations, totalRounds))
for iteration in range(num_iterations):
numPulls = np.zeros(numArms)
empReward = np.zeros(numArms)
regret = np.zeros(totalRounds)
for t in range(totalRounds):
#Initialise by pulling each arm once
if t < numArms:
numPulls[t] += 1
assert numPulls[t] == 1
reward = self.generate_sample(t)
empReward[t] = reward
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
continue
isCompetitive, _ = self.confidence_set_intersection(empReward, numPulls)
thompson = ThompsonSample(empReward,numPulls)
if isCompetitive.sum() == 0:
next_arm = =np.random.randint(0,self.numArms)
else:
self.next_arm_selection(thompson*isCompetitive)
#Generate reward, update pulls and empirical reward
reward_sample = self.generate_sample(next_arm)
empReward[next_arm] = (empReward[next_arm]*numPulls[next_arm] + reward_sample)/(numPulls[next_arm] + 1)
numPulls[next_arm] = numPulls[next_arm] + 1
#Evaluate regret
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
avg_regret[iteration, :] = regret
return avg_regret
def UCB_S(self, num_iterations, totalRounds):
avg_regret = np.zero((num_iterations, totalRounds))
for iteration in range(num_iterations):
numPulls = np.zeros(numArms)
empReward = np.zeros(numArms)
regret = np.zeros(totalRounds)
for t in range(totalRounds):
_, theta_hat = self.confidence_set_intersection(empReward, numPulls)
if len(theta_hat):
next_arm = np.random.randint(0,self.numArms)
else:
idx = np.array(list(theta_hat))
supReward = np.max(self.mu[idx, :], axis=0)
next_arm = self.next_arm_selection(supReward)
#Generate reward, update pulls and empirical reward
reward_sample = self.generate_sample(next_arm)
empReward[next_arm] = (empReward[next_arm]*numPulls[next_arm] + reward_sample)/(numPulls[next_arm] + 1)
numPulls[next_arm] = numPulls[next_arm] + 1
#Evaluate regret
regret[t] = mu_test[self.theta, bestArm] - mu_test[self.theta, next_arm]
avg_regret[iteration, :] = regret
return avg_regret
def run(self, num_iterations=20,T=5000):
avg_ucb_regret = self.UCB(num_iterations, T)
avg_ts_regret = self.TS(num_iterations, T)
avg_ucbc_regret = self.UCB_C(num_iterations, T)
avg_tsc_regret = self.TS_C(num_iterations, T)
avg_ucbs_regret = self.UCB_S(num_iterations, T)
# mean cumulative regret
self.plot_av_ucb = np.mean(avg_ucb_regret, axis=0)
self.plot_av_ts = np.mean(avg_ts_regret, axis=0)
self.plot_av_ucbc = np.mean(avg_ucbc_regret, axis=0)
self.plot_av_tsc = np.mean(avg_tsc_regret, axis=0)
self.plot_av_ucbs = np.mean(avg_ucbs_regret, axis=0)
# std dev over runs
self.plot_std_ucb = np.sqrt(np.var(avg_ucb_regret, axis=0))
self.plot_std_ts = np.sqrt(np.var(avg_ts_regret, axis=0))
self.plot_std_ucbc = np.sqrt(np.var(avg_ucbc_regret, axis=0))
self.plot_std_tsc = np.sqrt(np.var(avg_tsc_regret, axis=0))
self.plot_std_ucbs = np.sqrt(np.var(avg_ucbs_regret, axis=0))
self.save_data()
def save_data(self):
algorithms = ['ucb', 'ts', 'ucbc', 'tsc', 'ucbs']
pathlib.Path(f'data/plot_arrays/').mkdir(parents=False, exist_ok=True)
for alg in algorithms:
np.save(f'data/plot_arrays/plot_av_{alg}',
getattr(self, f'plot_av_{alg}'))
np.save(f'data/plot_arrays/plot_std_{alg}',
getattr(self, f'plot_std_{alg}'))
def plot(self):
spacing = 250
# Means
plt.plot(range(0, 5000)[::spacing], self.plot_av_ucb[::spacing], label='UCB', color='red', marker='+')
plt.plot(range(0, 5000)[::spacing], self.plot_av_ts[::spacing], label='TS', color='yellow', marker='o')
plt.plot(range(0, 5000)[::spacing], self.plot_av_ucbc[::spacing], label='UCB-C', color='blue', marker='^')
plt.plot(range(0, 5000)[::spacing], self.plot_av_tsc[::spacing], label='TS-C', color='black', marker='x')
plt.plot(range(0, 5000)[::spacing], self.plot_av_ucbs[::spacing], label='UCB-S', color='green', marker='*')
# Confidence bounds
plt.fill_between(range(0, 5000)[::spacing], (self.plot_av_ucb + self.plot_std_ucb)[::spacing],
(self.plot_av_ucb - self.plot_std_ucb)[::spacing], alpha=0.3, facecolor='r')
plt.fill_between(range(0, 5000)[::spacing], (self.plot_av_ts + self.plot_std_ts)[::spacing],
(self.plot_av_ts - self.plot_std_ts)[::spacing], alpha=0.3, facecolor='y')
plt.fill_between(range(0, 5000)[::spacing], (self.plot_av_ucbc + self.plot_std_ucbc)[::spacing],
(self.plot_av_ucbc - self.plot_std_ucbc)[::spacing], alpha=0.3, facecolor='b')
plt.fill_between(range(0, 5000)[::spacing], (self.plot_av_tsc + self.plot_std_tsc)[::spacing],
(self.plot_av_tsc - self.plot_std_tsc)[::spacing], alpha=0.3, facecolor='k')
plt.fill_between(range(0, 5000)[::spacing], (self.plot_av_ucbs + self.plot_std_ucbs)[::spacing],
(self.plot_av_ucbs - self.plot_std_ucbs)[::spacing], alpha=0.3, facecolor='g')
# Plot
plt.legend()
plt.grid(True, axis='y')
plt.xlabel('Number of Rounds')
plt.ylabel('Cumulative Regret')
# Save
pathlib.Path('data/plots/').mkdir(parents=False, exist_ok=True)
plt.savefig(f'data/plots/figure.pdf')
def parse_arguments():
parser = argparse.ArgumentParser()
parser.add_argument('--num_iterations', dest='num_iterations', type=int, default=20,
help="Number of iterations of each run")
parser.add_argument('--T', dest='T', type=int, default=5000, help="Number of rounds")
return parser.parse_args()
def main(args):
args = parse_arguments()
bandit_obj = algs()
bandit_obj.run(args.num_iterations, args.T)
bandit_obj.plot()
if __name__ == '__main__':
main(sys.argv)