-
Notifications
You must be signed in to change notification settings - Fork 34
/
mp_mo_q_learning.py
278 lines (252 loc) · 11.8 KB
/
mp_mo_q_learning.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
"""Outer-loop MOQ-learning algorithm (uses multiple weights)."""
import time
from copy import deepcopy
from typing import List, Optional
from typing_extensions import override
import gymnasium as gym
import numpy as np
from morl_baselines.common.evaluation import (
log_all_multi_policy_metrics,
policy_evaluation_mo,
)
from morl_baselines.common.morl_algorithm import MOAgent
from morl_baselines.common.scalarization import weighted_sum
from morl_baselines.common.weights import equally_spaced_weights, random_weights
from morl_baselines.multi_policy.linear_support.linear_support import LinearSupport
from morl_baselines.single_policy.ser.mo_q_learning import MOQLearning
class MPMOQLearning(MOAgent):
"""Multi-policy MOQ-Learning: Outer loop version of mo_q_learning.
Paper: Paper: K. Van Moffaert, M. Drugan, and A. Nowe, Scalarized Multi-Objective Reinforcement Learning: Novel Design Techniques. 2013. doi: 10.1109/ADPRL.2013.6615007.
"""
def __init__(
self,
env,
scalarization=weighted_sum,
learning_rate: float = 0.1,
gamma: float = 0.9,
initial_epsilon: float = 0.1,
final_epsilon: float = 0.1,
epsilon_decay_steps: int = None,
weight_selection_algo: str = "random",
epsilon_ols: Optional[float] = None,
use_gpi_policy: bool = False,
transfer_q_table: bool = True,
dyna: bool = False,
dyna_updates: int = 5,
gpi_pd: bool = False,
project_name: str = "MORL-Baselines",
experiment_name: str = "MultiPolicy MO Q-Learning",
wandb_entity: Optional[str] = None,
seed: Optional[int] = None,
log: bool = True,
):
"""Initialize the Multi-policy MOQ-learning algorithm.
Args:
env: The environment to learn from.
scalarization: The scalarization function to use.
learning_rate: The learning rate.
gamma: The discount factor.
initial_epsilon: The initial epsilon value.
final_epsilon: The final epsilon value.
epsilon_decay_steps: The number of steps for epsilon decay.
weight_selection_algo: The algorithm to use for weight selection. Options: "random", "ols", "gpi-ls"
epsilon_ols: The epsilon value for the optimistic linear support.
use_gpi_policy: Whether to use Generalized Policy Improvement (GPI) or not.
transfer_q_table: Whether to reuse a Q-table from a previous learned policy when initializing a new policy.
dyna: Whether to use Dyna-Q or not.
dyna_updates: The number of Dyna-Q updates to perform.
gpi_pd: Whether to use the GPI-PD method to prioritize Dyna updates.
project_name: The name of the project for logging.
experiment_name: The name of the experiment for logging.
wandb_entity: The entity to use for logging.
seed: The seed to use for reproducibility.
log: Whether to log or not.
"""
MOAgent.__init__(self, env, seed=seed)
# Learning
self.scalarization = scalarization
self.learning_rate = learning_rate
self.gamma = gamma
self.initial_epsilon = initial_epsilon
self.final_epsilon = final_epsilon
self.epsilon_decay_steps = epsilon_decay_steps
self.use_gpi_policy = use_gpi_policy
self.dyna = dyna
self.dyna_updates = dyna_updates
self.gpi_pd = gpi_pd
self.transfer_q_table = transfer_q_table
# Linear support
self.policies = []
self.weight_selection_algo = weight_selection_algo
self.epsilon_ols = epsilon_ols
assert self.weight_selection_algo in [
"random",
"ols",
"gpi-ls",
], f"Unknown weight selection algorithm: {self.weight_selection_algo}."
self.linear_support = LinearSupport(num_objectives=self.reward_dim, epsilon=epsilon_ols)
# Logging
self.project_name = project_name
self.experiment_name = experiment_name
self.log = log
if self.log:
self.setup_wandb(project_name=self.project_name, experiment_name=self.experiment_name, entity=wandb_entity)
@override
def get_config(self) -> dict:
return {
"env_id": self.env.unwrapped.spec.id,
"learning_rate": self.learning_rate,
"gamma": self.gamma,
"initial_epsilon": self.initial_epsilon,
"final_epsilon": self.final_epsilon,
"epsilon_decay_steps": self.epsilon_decay_steps,
"scalarization": self.scalarization.__name__,
"use_gpi_policy": self.use_gpi_policy,
"weight_selection_algo": self.weight_selection_algo,
"epsilon_ols": self.epsilon_ols,
"transfer_q_table": self.transfer_q_table,
"dyna": self.dyna,
"dyna_updates": self.dyna_updates,
"gpi_pd": self.gpi_pd,
"seed": self.seed,
}
def _gpi_action(self, state: np.ndarray, w: np.ndarray) -> int:
"""Get the action given by the GPI policy.
GPI(s, w) = argmax_a max_pi Q^pi(s, a, w) .
Args:
state: The state to get the action for.
weights: The weights to use for the scalarization.
Returns:
The action to take.
"""
q_vals = np.stack([policy.scalarized_q_values(state, w) for policy in self.policies])
_, action = np.unravel_index(np.argmax(q_vals), q_vals.shape)
return int(action)
def max_scalar_q_value(self, state: np.ndarray, w: np.ndarray) -> float:
"""Get the maximum Q-value over all policies for the given state and weights."""
return np.max([policy.scalarized_q_values(state, w) for policy in self.policies])
def eval(self, obs: np.array, w: Optional[np.ndarray] = None) -> int:
"""If use_gpi is True, return the action given by the GPI policy. Otherwise, chooses the best policy for w and follows it."""
if self.use_gpi_policy:
return self._gpi_action(obs, w)
else:
best_policy = np.argmax([np.dot(w, v) for v in self.linear_support.ccs])
return self.policies[best_policy].eval(obs, w)
def delete_policies(self, delete_indx: List[int]):
"""Delete the policies with the given indices."""
for i in sorted(delete_indx, reverse=True):
self.policies.pop(i)
def train(
self,
total_timesteps: int,
eval_env: gym.Env,
ref_point: np.ndarray,
known_pareto_front: Optional[List[np.ndarray]] = None,
timesteps_per_iteration: int = int(2e5),
num_eval_weights_for_front: int = 100,
num_eval_episodes_for_front: int = 5,
num_eval_weights_for_eval: int = 50,
eval_freq: int = 1000,
):
"""Learn a set of policies.
Args:
total_timesteps: The total number of timesteps to train for.
eval_env: The environment to use for evaluation.
ref_point: The reference point for the hypervolume calculation.
known_pareto_front: The optimal Pareto front, if known. Used for metrics.
timesteps_per_iteration: The number of timesteps per iteration.
num_eval_weights_for_front: The number of weights to use to construct a Pareto front for evaluation.
num_eval_episodes_for_front: The number of episodes to run when evaluating the policy.
num_eval_weights_for_eval (int): Number of weights use when evaluating the Pareto front, e.g., for computing expected utility.
eval_freq: The frequency of evaluation.
"""
if self.log:
self.register_additional_config(
{
"total_timesteps": total_timesteps,
"ref_point": ref_point.tolist(),
"known_front": known_pareto_front,
"timesteps_per_iteration": timesteps_per_iteration,
"num_eval_weights_for_front": num_eval_weights_for_front,
"num_eval_episodes_for_front": num_eval_episodes_for_front,
"num_eval_weights_for_eval": num_eval_weights_for_eval,
"eval_freq": eval_freq,
}
)
num_iterations = int(total_timesteps / timesteps_per_iteration)
if eval_env is None:
eval_env = deepcopy(self.env)
eval_weights = equally_spaced_weights(self.reward_dim, n=num_eval_weights_for_front)
for iter in range(num_iterations):
if self.weight_selection_algo == "ols" or self.weight_selection_algo == "gpi-ls":
w = self.linear_support.next_weight(
algo=self.weight_selection_algo,
gpi_agent=self if self.weight_selection_algo == "gpi-ls" else None,
env=eval_env if self.weight_selection_algo == "gpi-ls" else None,
rep_eval=num_eval_episodes_for_front,
)
if w is None:
print("OLS has no more corner weights to try. Using a random weight instead.")
w = random_weights(self.reward_dim, rng=self.np_random)
elif self.weight_selection_algo == "random":
w = random_weights(self.reward_dim, rng=self.np_random)
if len(self.policies) == 0 or not self.dyna:
model = None
else:
model = self.policies[-1].model # shared model
new_agent = MOQLearning(
env=self.env,
id=iter,
weights=w,
scalarization=self.scalarization,
learning_rate=self.learning_rate,
gamma=self.gamma,
initial_epsilon=self.initial_epsilon,
final_epsilon=self.final_epsilon,
epsilon_decay_steps=self.epsilon_decay_steps,
use_gpi_policy=self.use_gpi_policy,
dyna=self.dyna,
dyna_updates=self.dyna_updates,
model=model,
gpi_pd=self.gpi_pd,
parent=self,
log=self.log,
parent_rng=self.np_random,
seed=self.seed,
)
if self.transfer_q_table and len(self.policies) > 0:
reuse_ind = np.argmax([np.dot(w, v) for v in self.linear_support.ccs])
new_agent.q_table = deepcopy(self.policies[reuse_ind].q_table)
self.policies.append(new_agent)
start_time = time.time()
new_agent.global_step = self.global_step
new_agent.train(
start_time=start_time,
total_timesteps=timesteps_per_iteration,
reset_num_timesteps=False,
eval_freq=eval_freq,
eval_env=eval_env,
)
self.global_step = new_agent.global_step
value = policy_evaluation_mo(agent=new_agent, env=eval_env, w=w, rep=num_eval_episodes_for_front)[3]
removed_inds = self.linear_support.add_solution(value, w)
if self.weight_selection_algo != "random":
self.delete_policies(removed_inds)
if self.log:
if self.use_gpi_policy:
front = [
policy_evaluation_mo(agent=self, env=eval_env, w=w_eval, rep=num_eval_episodes_for_front)[3]
for w_eval in eval_weights
]
else:
front = self.linear_support.ccs
log_all_multi_policy_metrics(
current_front=front,
hv_ref_point=ref_point,
reward_dim=self.reward_dim,
global_step=self.global_step,
n_sample_weights=num_eval_weights_for_eval,
ref_front=known_pareto_front,
)
if self.log:
self.close_wandb()