def evaluate(self,
policy,
n_episodes=1,
initial_state=None,
transform_to_internal_state=None,
render=False,
results_dir='/tmp/'):
"""
Evaluate a policy on the mdp.
:param policy: a policy outputing an action
:type policy: Policy
:param n_episodes: number of episodes to evaluate
:type n_episodes: int
:param initial_state: the initial state to evaluate from
:type initial_state: int
:param transform_to_internal_state: transform the states or initial state to an internal state of the mdp
:type transform_to_internal_state: function
:param render: reder the environment
:type render: bool
:return: a Dataset with the trajectories
"""
dataset = Dataset(results_dir=results_dir)
for episode in range(n_episodes):
trajectory = []
state = self._env.reset()
if initial_state is not None:
state = initial_state
if self._quanser_robots:
self._env.env._sim_state = np.copy(
transform_to_internal_state(state))
else:
self._env.env.state = np.copy(
transform_to_internal_state(state))
for j in range(self._env._max_episode_steps):
with torch.no_grad():
state = torch.tensor(state,
device=policy.device,
dtype=TORCH_DTYPE)
action = policy(state).to('cpu').numpy().reshape((-1, ))
state_next, rew, done, _ = self._env.step(action)
trajectory.append((state.to('cpu').numpy(), action, rew,
state_next, done))
state = state_next
if render:
self._env.render()
if done:
break
dataset.add_trajectory(trajectory)
self._env.close()
return dataset
kwargs={
'fs': 200.0,
'fs_ctrl': 200.0
})
env = GentlyTerminating(gym.make('Qube-100-v1'))
dataset = Dataset()
n_trajectories = 1
for traj in range(n_trajectories):
trajectory = []
ctrl = SwingUpCtrl()
obs = env.reset()
done = False
while not done:
env.render()
act = ctrl(obs)
obs_n, r, done, _ = env.step(act)
trajectory.append((obs, act, r, obs_n, done))
if done:
break
obs = np.copy(obs_n)
env.close()
dataset.add_trajectory(trajectory)
dataset.update_dataset_internal()
print(np.sum(dataset._rewards))
filename = "../datasets/qube/{}_trajectories.npy".format(n_trajectories)
dataset.save_trajectories_to_file(filename)
