-
Notifications
You must be signed in to change notification settings - Fork 6
/
soft_randomMDPs_main.py
240 lines (206 loc) · 12.4 KB
/
soft_randomMDPs_main.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
import os
import sys
expname = sys.argv[1]
index = int(sys.argv[2])
import pandas as pd
import garnets
import spibb
import modelTransitions
from RMDP import *
from SPI import *
from shutil import copyfile
print('Start of experiment')
def safe_save(filename, df):
df.to_excel(filename + '.temp.xlsx')
copyfile(filename + '.temp.xlsx', filename + '.xlsx')
os.remove(filename + '.temp.xlsx')
print(str(len(results)) + ' lines saved to ' + filename + '.xlsx')
nb_trajectories_list = [10, 20, 50, 100, 200, 500, 1000, 2000]
delta = 1
epsilons = [0.1, 0.2, 0.5, 1, 2, 5]
ratios = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
seed = index
np.random.seed(seed)
gamma = 0.95
nb_states = 50
nb_actions = 4
nb_next_state_transition = 4
env_type = 2 # 1 for one terminal state, 2 for two terminal states
mask_0, thres = spibb.compute_mask(nb_states, nb_actions, 1, 1, [])
mask_0 = ~mask_0
rand_pi = np.ones((nb_states, nb_actions)) / nb_actions
filename = 'results/' + expname + '/results_' + str(index)
results = []
if not os.path.isdir('results'):
os.mkdir('results')
if not os.path.isdir('results/' + expname):
os.mkdir('results/' + expname)
self_transitions = 0
while True:
for ratio in ratios:
garnet = garnets.Garnets(nb_states, nb_actions, nb_next_state_transition,
env_type=env_type, self_transitions=self_transitions)
softmax_target_perf_ratio = (ratio + 1) / 2
baseline_target_perf_ratio = ratio
pi_b, q_pi_b, pi_star_perf, pi_b_perf, pi_rand_perf = \
garnet.generate_baseline_policy(gamma,
softmax_target_perf_ratio=softmax_target_perf_ratio,
baseline_target_perf_ratio=baseline_target_perf_ratio)
reward_current = garnet.compute_reward()
current_proba = garnet.transition_function
if env_type == 2: # easter
# Randomly pick a second terminal state and update model parameters
potential_final_states = [s for s in range(nb_states) if s != garnet.final_state and s != 0]
easter_egg = np.random.choice(potential_final_states)
# Or pick the one with the least transitions
# current_proba_sum = current_proba.reshape(-1, current_proba.shape[-1]).sum(axis=0)
# mask_easter = np.ma.array(current_proba_sum, mask=False)
# mask_easter.mask[garnet.final_state] = True
# easter_egg = np.argmin(mask_easter)
assert (garnet.final_state != easter_egg)
reward_current[:, easter_egg] = 1
current_proba[easter_egg, :, :] = 0
r_reshaped = spibb_utils.get_reward_model(current_proba, reward_current)
# Compute optimal policy in this new environment
true_rl = spibb.spibb(gamma, nb_states, nb_actions, mask_0, mask_0, current_proba, r_reshaped, 'default')
true_rl.fit()
pi_star_perf = spibb.policy_evaluation_exact(true_rl.pi, r_reshaped, current_proba, gamma)[0][0]
print("Optimal perf in easter egg environment:\t\t\t" + str(pi_star_perf))
pi_b_perf = spibb.policy_evaluation_exact(pi_b, r_reshaped, current_proba, gamma)[0][0]
print("Baseline perf in easter egg environment:\t\t\t" + str(pi_b_perf))
else:
easter_egg = None
r_reshaped = spibb_utils.get_reward_model(current_proba, reward_current)
for nb_trajectories in nb_trajectories_list:
# Generate trajectories, both stored as trajectories and (s,a,s',r) transition samples
trajectories, batch_traj = spibb_utils.generate_batch(nb_trajectories, garnet, pi_b, easter_egg)
print("GENERATED A DATASET OF " + str(nb_trajectories) + " TRAJECTORIES")
# Compute the maximal likelihood model for transitions and rewards.
# NB: the true reward function can be used for ease of implementation since it is not stochastic in our environment.
# One should compute it from the samples when it is stochastic.
model = modelTransitions.ModelTransitions(batch_traj, nb_states, nb_actions)
reward_model = spibb_utils.get_reward_model(model.transitions, reward_current)
# Estimates the values of the baseline policy with a monte-carlo estimation from the batch data:
# q_pib_est = spibb_utils.compute_q_pib_est(gamma, nb_states, nb_actions, trajectories)
# Computes the RL policy
rl = spibb.spibb(gamma, nb_states, nb_actions, pi_b, mask_0, model.transitions, reward_model, 'default')
rl.fit()
# Evaluates the RL policy performance
perfrl = spibb.policy_evaluation_exact(rl.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf RL:\t\t\t" + str(perfrl))
# Computes the Reward-adjusted MDP RL policy:
count_state_action = 0.00001 * np.ones((nb_states, nb_actions))
kappa = 0.003
for [action, state, next_state, reward] in batch_traj:
count_state_action[state, action] += 1
ramdp_reward_model = reward_model - kappa / np.sqrt(count_state_action)
ramdp = spibb.spibb(gamma, nb_states, nb_actions, pi_b, mask_0, model.transitions, ramdp_reward_model,
'default')
ramdp.fit()
# Evaluates the RL policy performance
perf_RaMDP = spibb.policy_evaluation_exact(ramdp.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf RaMDP:\t\t\t" + str(perf_RaMDP))
# Computes the Robust MDP policy:
terminal_state = 24
delta_RobustMDP = 0.001
rmdp = RMDP_based_alorithm(gamma, nb_states, nb_actions, delta_RobustMDP,
reward_current[0].reshape((nb_states, 1)), pi_b, terminal_state)
rmdp.fit(batch_traj)
safety_test = rmdp.safety_test()[0]
perf_RMDP_based_alorithm = spibb.policy_evaluation_exact(rmdp.pi_t, r_reshaped, current_proba, gamma)[0][0]
if safety_test:
perf_RMDP_based_alorithm_safe = perf_RMDP_based_alorithm
else:
perf_RMDP_based_alorithm_safe = pi_b_perf
print("perf RMDP_based_algorithm:\t" + str(perf_RMDP_based_alorithm))
# Computes the HCPI doubly robust policy:
delta_HCPI = 0.9
spi = SPI(gamma, pi_b, delta_HCPI, "student_t_test", "doubly_robust", trajectories, 0, 1, pi_b_perf,
reward_current)
pi_hcope = spi.get_policy()
perfhcope_doubly_robust = spibb.policy_evaluation_exact(pi_hcope, r_reshaped, current_proba, gamma)[0][0]
print("perf HCPI doubly_robust:\t" + str(perfhcope_doubly_robust))
N_wedge = 10
# Computation of the binary mask for the bootstrapped state actions
mask = spibb.compute_mask_N_wedge(nb_states, nb_actions, N_wedge, batch_traj)
## Policy-based SPIBB ##
# Computes the Pi_b_SPIBB policy:
pib_SPIBB = spibb.spibb(gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model,
'Pi_b_SPIBB')
pib_SPIBB.fit()
# Evaluates the Pi_b_SPIBB performance:
perf_Pi_b_SPIBB = spibb.policy_evaluation_exact(pib_SPIBB.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Pi_b_SPIBB:\t\t" + str(perf_Pi_b_SPIBB))
# Computes the Pi_<b_SPIBB policy:
pi_leq_b_SPIBB = spibb.spibb(gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model,
'Pi_leq_b_SPIBB')
pi_leq_b_SPIBB.fit()
# Evaluates the Pi_<b_SPIBB performance:
perf_Pi_leq_b_SPIBB = \
spibb.policy_evaluation_exact(pi_leq_b_SPIBB.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Pi_leq_b_SPIBB:\t\t" + str(perf_Pi_leq_b_SPIBB))
for epsilon in epsilons:
# Computation of the binary mask for the bootstrapped state actions
mask = spibb.compute_mask(nb_states, nb_actions, epsilon, delta, batch_traj)[0]
# Computation of the transition errors
errors = spibb.compute_errors(nb_states, nb_actions, delta, batch_traj)
## Soft-SPIBB 1-step ##
# Simplex (more variables than constraints)
soft_SPIBB_simplex_1step = spibb.spibb(
gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model, 'Soft_SPIBB_simplex',
errors=errors, epsilon=2 * epsilon, max_nb_it=1
)
soft_SPIBB_simplex_1step.fit()
# Evaluates the Soft-SPIBB-simplex performance
perf_soft_SPIBB_simplex_1step = \
spibb.policy_evaluation_exact(soft_SPIBB_simplex_1step.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Exact-Soft-SPIBB 1-step:\t" + str(perf_soft_SPIBB_simplex_1step))
# Computes the Soft-SPIBB-sort-Q policy
soft_SPIBB_sort_Q_1step = spibb.spibb(
gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model, 'Soft_SPIBB_sort_Q',
errors=errors, epsilon=2 * epsilon, max_nb_it=1
)
soft_SPIBB_sort_Q_1step.fit()
# Evaluates the Soft-SPIBB-sort-Q performance
perf_soft_SPIBB_sort_Q_1step = \
spibb.policy_evaluation_exact(soft_SPIBB_sort_Q_1step.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Approx-Soft-SPIBB 1-step:\t\t" + str(perf_soft_SPIBB_sort_Q_1step))
## Soft-SPIBB multi-steps ##
# Simplex (more variables than constraints)
soft_SPIBB_simplex = spibb.spibb(
gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model, 'Soft_SPIBB_simplex',
errors=errors, epsilon=2 * epsilon
)
soft_SPIBB_simplex.fit()
# Evaluates the Soft-SPIBB-simplex performance
perf_soft_SPIBB_simplex = \
spibb.policy_evaluation_exact(soft_SPIBB_simplex.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Exact-Soft-SPIBB:\t" + str(perf_soft_SPIBB_simplex))
# Computes the Soft-SPIBB-sort-Q policy
soft_SPIBB_sort_Q = spibb.spibb(
gamma, nb_states, nb_actions, pi_b, mask, model.transitions, reward_model, 'Soft_SPIBB_sort_Q',
errors=errors, epsilon=2 * epsilon
)
soft_SPIBB_sort_Q.fit()
# Evaluates the Soft-SPIBB-sort-Q performance
perf_soft_SPIBB_sort_Q = \
spibb.policy_evaluation_exact(soft_SPIBB_sort_Q.pi, r_reshaped, current_proba, gamma)[0][0]
print("perf Approx-Soft-SPIBB:\t\t" + str(perf_soft_SPIBB_sort_Q))
results.append([seed, gamma, nb_states, nb_actions, 4,
nb_trajectories, softmax_target_perf_ratio, baseline_target_perf_ratio,
pi_b_perf, pi_rand_perf, pi_star_perf, perfrl, perf_RaMDP, perf_RMDP_based_alorithm,
perfhcope_doubly_robust, perf_Pi_b_SPIBB, perf_Pi_leq_b_SPIBB,
perf_soft_SPIBB_simplex, perf_soft_SPIBB_sort_Q,
perf_soft_SPIBB_simplex_1step, perf_soft_SPIBB_sort_Q_1step,
kappa, delta_RobustMDP, delta_HCPI, N_wedge, epsilon
])
df = pd.DataFrame(results, columns=['seed', 'gamma', 'nb_states', 'nb_actions', 'nb_next_state_transition',
'nb_trajectories', 'softmax_target_perf_ratio', 'baseline_target_perf_ratio',
'baseline_perf', 'pi_rand_perf', 'pi_star_perf', 'perfrl',
'perf_RaMDP', 'perf_RMDP_based_algorithm', 'perfhcope_doubly_robust',
'perf_Pi_b_SPIBB', 'perf_Pi_leq_b_SPIBB',
'perf_soft_SPIBB_simplex', 'perf_soft_SPIBB_sort_Q',
'perf_soft_SPIBB_simplex_1step', 'perf_soft_SPIBB_sort_Q_1step',
'kappa', 'delta_RobustMDP', 'delta_HCPI', 'N_wedge', 'epsilon'])
# Save it to an excel file
safe_save(filename, df)