-
Notifications
You must be signed in to change notification settings - Fork 5
/
_Ohio_Simulator.py
341 lines (301 loc) · 15.6 KB
/
_Ohio_Simulator.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
TODO: allow online interating. take a step.
"""
""" Usage
1. evaluation
1. (multi-seeds) simulate a dataset with the behaviour policy, OPE the target policy.
2. evaluate in the simulator (multi-seeds)
2. offline RL
1. (multi-seeds) simulate a dataset with the behaviour policy, learn a policy (ours or competing), and evaluate in the simulator (multi-seeds)
"""
# import os, sys
# package_path = os.path.dirname(os.path.abspath(os.getcwd()))
# sys.path.insert(0, package_path + "/test_func")
# from _util_TRPO import *
from _util import *
import operator
import os
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["OPENBLAS_NUM_THREADS"] = "1"
########################################################################################################################################################################################################################################################################################################################################################################################
class OhioSimulator():
def __init__(self, sd_G = 3, T = 20, N = 10, T_burnin = 1000
, behav = None
, lag = 4, noiseless = False, equal_A = False):
# the following parameters will not change with the LM fitting
self.lag = lag
self.noiseless = noiseless
self.behav = behav
######
# if self.lag == 1:
# self.CONST = 0
# else:
self.CONST = 39.03
####################
self.init_u_G = 162
self.init_sd_G = 60
####################
self.p_D, self.u_D, self.sd_D = 0.17, 44.4, 35.5
self.p_E, self.u_E, self.sd_E = 0.05, 4.9, 1.04
self.range_a = [0, 1, 2, 3, 4]
# left to right: t-4, .. , t-1
if equal_A:
self.p_A = [0.2, 0.2, 0.2, 0.2, 0.2]
else:
self.p_A = [0.805, 0.084, 0.072, 0.029, 0.010] # new discritization
if self.lag == 1:
## debug only
if self.noiseless:
self.coefficients = [.8, 0, 0, -8]
#self.coefficients = [.8, 0, 0, -4]
else:
self.coefficients = [.8, 0.23, -3.489, 0]
elif self.lag == 4:
if self.noiseless:
self.coefficients = [-0.008 , 0.106 , -0.481 , 1.171 # glucose
, 0.00 , -0.00 , 0.0 , 0 # diet
, 0.00 , 0 , 0 , 0 # exercise
, -0.30402253, -2.02343638, -0.3310525 , -0.43941028] # action
else:
self.coefficients = [-0.008 , 0.106 , -0.481 , 1.171 # glucose
, 0.008 , -0.004 , 0.08 , 0.23 # diet
, 0.009 , -1.542 , 3.097 , -3.489 # exercise
#, 0, 0, 0, 0]
, -0.30402253, -2.02343638, -0.3310525 , -0.43941028] # action
self.tran_mat = np.expand_dims(arr(self.coefficients), 0)
####################
self.sd_G = sd_G
self.T, self.N = T, N
self.seed = 42
self.T_burnin = T_burnin
def Glucose2Reward(self, gls):
low_gl, high_gl = 80, 140
rewards = np.select([gls >= high_gl, gls <= low_gl, np.multiply(low_gl < gls, gls < high_gl)]
, [-(gls - high_gl) ** 1.35 / 30, -(low_gl - gls) ** 2 / 30, 0])
return rewards
############################################################################################################################################
############################################################################################################################################
def init_MDPs(self, seed = 0, N = None):
""" Randomly initialize
1. G_t [0,..., 4]
1. the other state variable: random.
2. errors for G_t
3. when to take how many diets/exercises [matters?]
where T varies, seed is diff; dominated by init several states, and hence values are not monotone.
self.T_burnin should only be here
"""
np.random.seed(seed)
if N is None:
N = self.N
T = self.T + self.T_burnin
obs = np.zeros((3, T, N)) # [Gi, D, Ex]
e_D = abs(rnorm(self.u_D, self.sd_D, T * N))
e_E = abs(rnorm(self.u_E, self.sd_E, T * N))
e_G = rnorm(0, self.sd_G, T * N).reshape((T, N))
obs[0, :self.lag, :] = rnorm(self.init_u_G, self.init_sd_G, self.lag * N).reshape(self.lag, N)
obs[1, :, :] = (rbin(1, self.p_D, T * N) * e_D).reshape((T, N))
obs[2, :, :] = (rbin(1, self.p_E, T * N) * e_E).reshape((T, N))
actions = np.random.choice(range(len(self.p_A)), size = T * N, p = self.p_A).reshape((T, N))
######### Transition
for t in range(self.lag - 1, self.T_burnin + 5):
if self.behav is not None:
S = self.conc_SA_2_state(obs, actions, t, multiple_N = True).T
actions[t, :] = np.squeeze(self.behav.sample_A(S))
states = self.concatenate_useful_obs(obs = obs, actions = actions, t = t + 1)
obs[0, t + 1, :] = self.step(states = states, errors = np.array([e_G[t, :]]))
actions = actions.astype(float)
return obs[:, self.T_burnin:, :], e_G[self.T_burnin:, :], actions[self.T_burnin:, :]
##################################################################################################################################################################
def conc_single(self, obs, actions, t):
# obs = [3, T]
# actions = [T]
As = actions[(t - self.lag + 1):t]
S = obs[:, (t - self.lag + 1):(t + 1)]
s = S.ravel(order='C')
s = np.append(s, As)
return s
############################################################################################################################################
def step(self, states = None, errors = None):
return np.array(self.CONST).reshape((1, 1)) + self.tran_mat.dot(states) + errors
def simu_one_seed(self, seed = 42, N = None, T = None):
""" Simulate N patient trajectories with length T, calibrated from the Ohio dataset.
Returns:
trajs = [traj], where traj [[S, A, R, SS]]
"""
if N is None:
N = self.N
if T is None:
T = self.T
np.random.seed(seed)
# Initialization
obs, e_G, actions = self.init_MDPs(seed = seed, N = N)
for t in range(self.lag - 1, T - 1):
if self.behav is not None:
S = self.conc_SA_2_state(obs, actions, t, multiple_N = True).T
actions[t, :] = np.squeeze(self.behav.sample_A(S))
states = self.concatenate_useful_obs(obs = obs, actions = actions, t = t + 1)
obs[0, t + 1, :] = self.step(states = states, errors = np.array([e_G[t, :]]))
actions = actions.astype(float)
######### Collection: obs = (3, T, N), actions = (T, N)
# conc_SA_2_state(obs, actions, t, multiple_N = False, J = 4)
trajs = [[] for i in range(N)]
for t in range(self.lag - 1, T - 1): # -1, 10/01/2020
conc_s = self.conc_SA_2_state(obs, actions, t, multiple_N = True) # (dim_with_J, N)
conc_ss = self.conc_SA_2_state(obs, actions, t + 1, multiple_N = True)# (dim_with_J, N)
As = actions[t, :]
Rs = self.Glucose2Reward(conc_ss[-3, :])
for i in range(N):
trajs[i].append([conc_s[:, i], As[i], Rs[i], conc_ss[:, i]])
return trajs
def simu_init_S(self, seed = 42, N = None):
if N is None:
N = self.N
trajs = self.simu_one_seed(seed = seed, N = N, T = 6)
return arr([trajs[i][0][0] for i in range(N)])
############################################################################################################################################
############################################################################################################################################
def eval_policy(self, pi = None, N = None
, gamma = 1, seed = 42, return_init = False, return_value = True, return_init_value = False
, init_S = None, init_A = None
):
""" Evaluate the value of a policy in simulation.
sample the first four time points，and then begin to follow the policy and collect rewards
transform into matrix so that linear transition is easier
Concatenate data for training and evaluation
"""
pi.seed = seed
np.random.seed(seed)
T = self.T
if N is None:
N = self.N
##############################
## after burn-in
obs, e_G, actions_init = self.init_MDPs(seed = seed, N = N) # obs = [3, T, N]
self.e_G = e_G
actions = np.zeros((T, N)) # store previous actions+
actions[:(self.lag - 1), :] = actions_init[:(self.lag - 1), :]
if init_S is not None:
obs[:, :self.lag, :] = init_S
actions[:(self.lag - 1), :] = init_A
# self.init = [obs.copy(), e_G.copy(), actions[:self.lag, :].copy()]
##############################
curr_time = now()
if return_init:
# can be used for references?
return obs[:, :self.lag, :], actions[:(self.lag - 1), :]
for t in range(self.lag - 1, T - 1):
# choose actions based on status. obs = [3, T, N]
S = self.conc_SA_2_state(obs, actions, t, multiple_N = True).T # N * dim
actions[t, :] = np.squeeze(pi.sample_A(S)) # s [N * dx] -> actions [N * 1] -> 1 * N
# next observations: based on ..., t-1, to decide t.
states = self.concatenate_useful_obs(obs = obs, actions = actions, t = t + 1)
obs[0, t + 1, :] = self.step(states = states, errors = np.array([e_G[t, :]]))
##############################
if return_value:
discounted_values, average_values = self.cal_reward(obs, gamma)
if return_init_value:
return discounted_values
printR("True Value: mean = {:.2f} with std = {:.2f}".format(np.mean(discounted_values), np.std(discounted_values) / np.sqrt(N)))
return mean(discounted_values) # len-N
else:
######### Collection: obs = (3, T, N), actions = (T, N)
trajs = [[] for i in range(N)]
for t in range(self.lag - 1, T - 1):
conc_s = self.conc_SA_2_state(obs, actions, t, multiple_N = True) # (dim_with_J, N)
conc_ss = self.conc_SA_2_state(obs, actions, t + 1, multiple_N = True)# (dim_with_J, N)
As = actions[t, :]
Rs = self.Glucose2Reward(conc_ss[-3, :])
for i in range(N):
trajs[i].append([conc_s[:, i], As[i], Rs[i], conc_ss[:, i]])
return trajs
def cal_reward(self, obs, gamma):
all_rewards = self.Glucose2Reward(obs)
all_rewards = all_rewards[0]
all_rewards = all_rewards[self.lag:]
#all_rewards = np.roll(all_rewards[self.lag:], shift = -1, axis = 0)
all_rewards = np.squeeze(all_rewards)
#gammas = np.expand_dims(arr([gamma ** j for j in range(all_rewards.shape[0])]), 0)
discounted_values = sum(r * gamma ** j for j, r in enumerate(all_rewards))
#discounted_values = np.dot(gammas, all_rewards)
average_values = np.mean(all_rewards) # 0
return discounted_values, average_values
############################################################################################################################################
############################################################################################################################################
def concatenate_useful_obs(self, obs, actions, t):
# (dim_with_J, N)
r = np.vstack([
obs[0, (t - self.lag):t, :], obs[1, (t - self.lag):t, :],
obs[2, (t - self.lag):t, :], actions[(t - self.lag):t, :]])
return r
def conc_SA_2_state(self, obs, actions, t, multiple_N = False):
""" to form a lag-J states from history obs and A
"""
# dim = (3, T, N)
N = obs.shape[2]
dim_obs = 3
s = np.vstack(([
obs[:, (t - self.lag + 1 ):t, :],
actions[(t - self.lag + 1):t, :].reshape((1, self.lag - 1, N))])) # extend_dim for first one
s = s.reshape(((dim_obs + 1) * (self.lag - 1), N), order = 'F')
obs_0 = obs[:, t, :] # 3 * N
s = np.vstack([s, obs_0])
return s
############################################################################################################################################
############################################################################################################################################
def simu_multi_seeds(self, M, parallel = True):
if parallel:
return parmap(self.simu_one_seed, range(M))
else:
return [self.simu_one_seed(i) for i in range(M)]
def eval_trajs(self, trajs = None, gamma = .8):
def get_rew(i):
glucoses = arr([item[0][-3] for item in trajs[i]])
rewards = np.roll(self.Glucose2Reward(glucoses), shift = -1).reshape(-1, 1)
discounted_value = sum(r * gamma ** t for t, r in enumerate(rewards))
discounted_value = round(discounted_value, 2)
average_value = round(np.mean(rewards))
return [rewards, discounted_value, average_value]
res = parmap(get_rew, range(len(trajs)))
discounted_values = [a[1] for a in res]
average_values = [a[2] for a in res]
return discounted_values, average_values
############################################################################################################################################
############################################################################################################################################
def reset(self, T = 1000):
self.seed += 1
if self.seed > 1e4:
self.seed = 42
self.T = T
np.random.seed(self.seed)
### Initialization
self.obs = np.zeros((3, T))
self.e_D = abs(rnorm(self.u_D, self.sd_D, T))
self.e_E = abs(rnorm(self.u_E, self.sd_E, T))
self.e_G = rnorm(0, self.sd_G, T)
self.obs[0, :self.lag] = rnorm(self.init_u_G, self.init_sd_G, self.lag)
self.obs[1, :] = rbin(1, self.p_D, T) * self.e_D
self.obs[2, :] = rbin(1, self.p_E, T) * self.e_E
self.t = 3
self.actions = zeros(T)
self.actions[:3] = np.zeros(3)
states = self.conc_single(obs = self.obs, actions = self.actions, t = self.t)
return states
def online_step(self, action):
self.actions[self.t] = action
# 1. transition
states = self.conc_single(obs = self.obs, actions = self.actions, t = self.t)
SA = np.append(states, action)
self.t += 1
next_G = self.CONST + self.tran_mat.dot(SA) + self.e_G[self.t]
# 2. store
self.obs[0, self.t] = next_G
# 3. return
observation_ = self.conc_single(obs = self.obs, actions = self.actions, t = self.t)
done = (self.t == (self.T - 1))
reward = self.Glucose2Reward(next_G)
reward += (randn(1) / 1e4)
reward = np.squeeze(reward)
return observation_, reward, done
############################################################################################################################################