Port TRPO, PPO, GAIL

scratch/etienne/intersimple/trpo_speed_lidar.py

@@ -0,0 +1,52 @@
+# %%
+from sb3_contrib import TRPO
+from intersim.envs import IntersimpleLidarFlat
+from intersim.envs.intersimple import speed_reward
+import functools
+
+model_name = "trpo_speed_lidar"
+
+#def reward(state, action, info):
+#    speed = state[2].item()
+#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
+#    return 0.1 * r
+
+env = IntersimpleLidarFlat(
+    n_rays=5,
+    agent=51,
+    reward=functools.partial(
+        speed_reward,
+        collision_penalty=0
+    ),
+)
+
+# %%
+model = TRPO(
+    "MlpPolicy", env,
+    learning_rate=1e-4,
+    verbose=1,
+    tensorboard_log='runs/',
+    #use_sde=True,
+    #sde_sample_freq=4,
+)
+model.learn(total_timesteps=1000000)
+model.save(model_name)
+
+print('Done training.')
+
+del model # remove to demonstrate saving and loading
+
+# %%
+model = TRPO.load(model_name)
+
+obs = env.reset()
+while True:
+    action, _states = model.predict(obs)
+    obs, rewards, done, info = env.step(action)
+    env.render(mode='post')
+    if done:
+        break
+
+env.close(filestr='render/'+model_name)
+
+# %%