-
Notifications
You must be signed in to change notification settings - Fork 0
/
bandit_environment.py
184 lines (153 loc) · 7.09 KB
/
bandit_environment.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
import numpy as np
import gym
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import os
from sklearn.cluster import KMeans
from scipy.spatial.distance import pdist
from scipy.stats import pearsonr
from synthetic_gaussian_mapping import SyntheticGaussianMapping
np.set_printoptions(precision=4)
gym.logger.set_level(40)
sns.set_theme(style='darkgrid', palette='muted', font='monospace')
class HyperpersonalizationEnv:
'''generate contextual bandit for a synthetic hyperpersonalization task'''
def __init__(self, params_env):
self.__dict__.update(params_env)
self.set_random_seed()
self.sample_actions()
self.create_reward_function()
def set_random_seed(self, seed=None):
'''fix random seed for reproducibility'''
if seed is not None:
self.seed = seed
np.random.seed(seed)
self.rng_s = np.random.default_rng(seed=self.seed)
self.rng_a = np.random.default_rng(seed=self.seed+1)
self.rng_r = np.random.default_rng(seed=self.seed+2)
def get_state(self, num_s=1):
'''generate observed states'''
S = self.rng_s.uniform(self.s_low, self.s_high, (num_s,self.dim_s))
return S
def sample_actions(self):
'''generate the set of available actions'''
self.A = self.rng_a.uniform(self.a_low, self.a_high, (self.num_a,self.dim_a))
def create_state_embedding(self):
'''generate state feature map'''
self.params_s = {'dim_in': self.dim_s, 'dim_layers': self.r_arch,
'dim_out': self.dim_feature, 'seed': self.rng_r.integers(1e+09)}
self.feature_map_s = SyntheticGaussianMapping(self.params_s)
self.feature_s = lambda s: self.feature_map_s.propagate(s)
def create_action_embedding(self):
'''generate action feature map'''
self.params_a = {'dim_in': self.dim_a, 'dim_layers': self.r_arch,
'dim_out': self.dim_feature, 'seed': self.rng_r.integers(1e+09)}
self.feature_map_a = SyntheticGaussianMapping(self.params_a)
self.feature_a = lambda a: self.feature_map_a.propagate(a)
def feature_relevance(self, s, a):
'''measure feature relevance in the latent feature space'''
norm_s = np.linalg.norm(s, axis=1, keepdims=True)
norm_a = np.linalg.norm(a, axis=1, keepdims=True)
cosine_sim = np.matmul(s, a.T) / np.matmul(norm_s, norm_a.T)
return cosine_sim
def create_reward_function(self):
'''reward function of the environment'''
self.create_state_embedding()
self.create_action_embedding()
self.r = lambda s,a: self.feature_relevance(self.feature_s(s), self.feature_a(a))
def get_r_vals(self, s):
'''compute the average, minimum, and maximum reward values for a given state'''
r_vals = self.r(s, self.A)
r_avg = r_vals.mean(axis=1)
r_min = r_vals.min(axis=1)
r_max = r_vals.max(axis=1)
return r_avg, r_min, r_max
def visualize_reward_distribution(self, num_s=5, num_a=10):
'''visualize reward distribution across the available actions'''
S = np.random.randn(num_s, self.dim_s) / 100
R = pd.DataFrame(self.r(S, self.A[:num_a]).T)
ax = R.plot.bar(figsize=(8,5), width=.8, rot=0, legend=None)
ax.set_xlabel('available actions')
ax.set_ylabel('reward')
os.makedirs('./images/', exist_ok=True)
plt.savefig('./images/reward_distribution.pdf', format='pdf')
plt.show()
def visualize_reward_correlation(self, num_s=100000):
'''visualize reward correlations across different clusters'''
S = self.get_state(num_s=num_s)
R = self.r(S, self.A)
clusters = KMeans(n_clusters=self.num_a).fit(S).labels_
corr = np.zeros(shape=(2,self.num_a))
for k in range(self.num_a):
k_ind = np.where(clusters==k)[0]
corr[:,k] = pearsonr(pdist(S[k_ind]), pdist(R[k_ind]))
fig, ax = plt.subplots(figsize=(8,5))
ax.bar(np.arange(1,self.num_a+1), corr[0], width=1.)
ax.set_title(f'Pearson correlation (p = {np.mean(corr[1]):.2e})')
ax.set_xlabel('state space clusters')
os.makedirs('./images/', exist_ok=True)
plt.savefig('./images/reward_correlation.pdf', format='pdf')
plt.show()
class GymHyperpersonalizationEnv(gym.Env):
'''create custom gym environment for a hyperpersonalization task'''
def __init__(self, params_env):
super(GymHyperpersonalizationEnv, self).__init__()
self.env = HyperpersonalizationEnv(params_env)
self.action_space = gym.spaces.Discrete(self.env.num_a)
self.observation_space = gym.spaces.Box(low=self.env.s_low, high=self.env.s_high,
shape=(self.env.dim_s,), dtype=np.float)
def reset_env(self):
'''re-seed the environment'''
self.env.set_random_seed()
def step(self, action_index):
'''given an observed state take an action and receive reward'''
self.action = self.env.A[action_index]
self.reward = self.env.r(self.state, self.action).item()
done = True
info = {}
return self.state, self.reward, done, info
def reset(self):
'''observe a new state'''
self.state = self.env.get_state().flatten()
return self.state
def get_r_stats(self, steps):
'''compute the average, minimum, and maximum reward values'''
S = self.env.get_state(steps)
stats = self.env.get_r_vals(S)
self.reset_env()
return stats
class GymStateClusteringEnv(GymHyperpersonalizationEnv):
'''create custom gym environment with a clustered state space'''
def __init__(self, params_env, params_cluster):
GymHyperpersonalizationEnv.__init__(self, params_env)
self.__dict__.update(params_cluster)
self.action_space = gym.spaces.Discrete(self.env.num_a)
self.observation_space = gym.spaces.Discrete(self.K)
self.cluster_state_space()
def cluster_state_space(self):
'''cluster the state space'''
S = self.env.get_state(self.timesteps)
self.clustering = KMeans(n_clusters=self.K).fit(S)
self.reset_env()
def step(self, action_index):
'''given an observed state take an action and receive reward'''
self.action = self.env.A[action_index]
self.reward = self.env.r(self.state_vector, self.action).item()
done = True
info = {}
return self.state, self.reward, done, info
def reset(self):
'''observe a new state'''
self.state_vector = self.env.get_state()
self.state = self.clustering.predict(self.state_vector).item()
return self.state
if __name__ == '__main__':
'''configure the environment'''
params_env = {'num_a': 100, 'dim_s': 100, 'dim_a': 100, 'dim_feature': 100,
's_low': -1, 's_high': 1, 'a_low': -1, 'a_high': 1,
'r_arch': [100,100,100], 'seed': 2021}
'''create the environment'''
env = HyperpersonalizationEnv(params_env)
env.visualize_reward_distribution()
env.visualize_reward_correlation()