-
Notifications
You must be signed in to change notification settings - Fork 0
/
training_baseline_policy.py
174 lines (159 loc) · 5.33 KB
/
training_baseline_policy.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
import argparse
import pickle
import os
import matplotlib.pyplot as plt
import pandas
import seaborn
import yaml
import pogym
from agents import FSCAgent
from agents import QLearningFSCAgent
from fsc import FiniteObservationHistoryFSC
from utils import evaluate_agent
from utils.training import run
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env_id",
default="Tiger-v0",
type=str,
choices=pogym.env_ids,
help="gym env id (default=Tiger-v0)"
)
parser.add_argument(
"--training_episodes",
default=5000,
type=int,
help="number of training episodes (default=5000)"
)
parser.add_argument(
"--evaluation_episodes",
default=10000,
type=int,
help="number of evaluation episodes (default=10000)"
)
parser.add_argument(
"--decaying_rate_qlearning",
default=0.002,
type=float,
help="q-learning exponential decay rate (default=0.002)"
)
parser.add_argument(
"--beta",
default=0.002,
type=float,
help="softmax temperature to generate the behavior policy (default=0.002)"
)
parser.add_argument(
"--k",
default=1,
type=int,
help="size of history the FSC is tracking (default=1)"
)
parser.add_argument(
"--ma_size",
default=0.05,
type=float,
help="size of window for moving average -- percentage of the number of training episodes, (default=0.05)"
)
parser.add_argument(
"--seed",
default=123,
type=int,
help="seed (default=123)"
)
parser.add_argument(
"--discount",
default=0.95,
type=float,
help="discount factor -- gamma (default=0.95)"
)
parser.add_argument(
"--out_dir",
default="data",
type=str,
help="output directory (default=data)"
)
return parser.parse_args()
def train_agent_on_finite_history(seed, training_episodes, env_id, decaying_rate_qlearning, beta, discount, ma_size,
evaluation_episodes, k, out_dir):
assert 0 < ma_size < 1
window = int(training_episodes * ma_size) # size of the moving average window
training_episodes += window
agent_label = f"Q-Learning FSC({k})"
os.makedirs(out_dir, exist_ok=True)
output_prefix = os.path.join(out_dir, f"{env_id}_k-{k}")
env = pogym.make(env_id)
fsc = FiniteObservationHistoryFSC.make_uniform_fsc(env.observation_space.n, env.action_space.n, k=k)
env.reset(seed=seed)
agent = QLearningFSCAgent(env, seed=seed, fsc=fsc, alpha=1, discount=discount, epsilon=0.5,
decaying_rate=decaying_rate_qlearning)
episodic_returns = run(agent, env, training_episodes, discount, verbose=True, label=agent_label)
final_agent = FSCAgent(env, agent.export_fsc(), seed)
final_performance = evaluate_agent(final_agent, env, num_episodes=evaluation_episodes, disc=discount, verbose=True)
baseline_fsc = agent.export_fsc(beta)
baseline_agent = FSCAgent(env, baseline_fsc, seed)
baseline_performance = evaluate_agent(baseline_agent, env, num_episodes=evaluation_episodes, disc=discount, verbose=True)
with open(f'{output_prefix}.pkl', 'wb') as f:
pickle.dump({
"baseline_fsc": baseline_fsc,
"baseline_performance": baseline_performance,
"final_agent": final_agent,
"final_performance": final_performance,
"training_episodes": training_episodes,
"env_id": env_id,
"k": k,
"decaying_rate_qlearning": decaying_rate_qlearning,
"seed": seed,
"beta": beta,
"discount": discount,
}, f)
with open(f'{output_prefix}.yaml', 'w') as f:
yaml.dump({
"baseline_fsc": str(baseline_fsc),
"baseline_performance": baseline_performance,
"final_agent": str(final_agent),
"final_performance": final_performance,
"training_episodes": training_episodes,
"env_id": env_id,
"k": k,
"decaying_rate_qlearning": decaying_rate_qlearning,
"seed": seed,
"beta": beta,
"discount": discount,
}, f)
df = pandas.DataFrame.from_dict(
dict(
rewards=episodic_returns,
episode=range(training_episodes),
agent=agent_label,
seed=seed,
)
)
df['rewards_ma'] = df["rewards"].rolling(window, min_periods=window).mean().shift(-window)
ax = seaborn.lineplot(data=df, x="episode", y="rewards_ma", hue="agent", ci=None, legend=True)
seaborn.lineplot(
data=df, x="episode", y=final_performance,
ci=None,
ax=ax,
linestyle="--",
legend=True,
label="Final Policy"
)
seaborn.lineplot(
data=df, x="episode", y=baseline_performance,
ci=None,
ax=ax,
linestyle="-.",
legend=True,
label="Behavior Policy"
)
ax.set_xlabel("Episode")
ax.set_ylabel(f"Return (moving average {window})")
ax.set_title(f'{env_id}')
plt.savefig(f"{output_prefix}.pdf")
plt.savefig(f"{output_prefix}.png")
plt.show()
plt.clf()
if __name__ == '__main__':
train_agent_on_finite_history(**vars(parse_args()))