Add code
Browse files
README.md
CHANGED
@@ -24,5 +24,59 @@ model-index:
|
|
24 |
This is a trained model of a **DQN** agent playing **LunarLander-v2** using the [stable-baselines3 library](https://github.com/DLR-RM/stable-baselines3).
|
25 |
|
26 |
## Usage (with Stable-baselines3)
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
|
|
24 |
This is a trained model of a **DQN** agent playing **LunarLander-v2** using the [stable-baselines3 library](https://github.com/DLR-RM/stable-baselines3).
|
25 |
|
26 |
## Usage (with Stable-baselines3)
|
27 |
+
|
28 |
+
```python
|
29 |
+
from stable_baselines3 import DQN
|
30 |
+
from stable_baselines3.common.env_util import make_vec_env
|
31 |
+
from stable_baselines3.common.callbacks import EvalCallback
|
32 |
+
|
33 |
+
# Create the environment
|
34 |
+
env_id = "LunarLander-v2"
|
35 |
+
n_envs = 1
|
36 |
+
env = make_vec_env(env_id, n_envs=n_envs)
|
37 |
+
|
38 |
+
# Create the evaluation envs
|
39 |
+
eval_envs = make_vec_env(env_id, n_envs=5)
|
40 |
+
|
41 |
+
# Adjust evaluation interval depending on the number of envs
|
42 |
+
eval_freq = int(1e5)
|
43 |
+
eval_freq = max(eval_freq // n_envs, 1)
|
44 |
+
|
45 |
+
# Create evaluation callback to save best model
|
46 |
+
# and monitor agent performance
|
47 |
+
eval_callback = EvalCallback(
|
48 |
+
eval_envs,
|
49 |
+
best_model_save_path="./logs/",
|
50 |
+
eval_freq=eval_freq,
|
51 |
+
n_eval_episodes=10,
|
52 |
+
)
|
53 |
+
|
54 |
+
# Instantiate the agent
|
55 |
+
# Hyperparameters from https://github.com/DLR-RM/rl-baselines3-zoo
|
56 |
+
model = DQN(
|
57 |
+
"MlpPolicy",
|
58 |
+
env,
|
59 |
+
learning_starts=0,
|
60 |
+
batch_size=128,
|
61 |
+
buffer_size=50000,
|
62 |
+
learning_rate=1e-3,
|
63 |
+
target_update_interval=250,
|
64 |
+
train_freq=4,
|
65 |
+
gradient_steps=-1,
|
66 |
+
# Explore for 20_000 timesteps
|
67 |
+
exploration_fraction=0.04,
|
68 |
+
exploration_final_eps=0.1,
|
69 |
+
policy_kwargs=dict(net_arch=[256, 256]),
|
70 |
+
verbose=1,
|
71 |
+
)
|
72 |
+
|
73 |
+
# Train the agent (you can kill it before using ctrl+c)
|
74 |
+
try:
|
75 |
+
model.learn(total_timesteps=int(5e5), callback=eval_callback)
|
76 |
+
except KeyboardInterrupt:
|
77 |
+
pass
|
78 |
+
|
79 |
+
# Load best model
|
80 |
+
model = DQN.load("logs/best_model.zip")
|
81 |
+
```
|
82 |
|