def test_mbmpo_compilation(self):
"""Test whether an MBMPOTrainer can be built with all frameworks."""
config = (mbmpo.MBMPOConfig().rollouts(
num_rollout_workers=2, horizon=200).training(dynamics_model={
"ensemble_size": 2
}).environment(
env="ray.rllib.examples.env.mbmpo_env.CartPoleWrapper"))
num_iterations = 1
# Test for torch framework (tf not implemented yet).
for _ in framework_iterator(config, frameworks="torch"):
trainer = config.build()
for i in range(num_iterations):
results = trainer.train()
check_train_results(results)
print(results)
check_compute_single_action(trainer,
include_prev_action_reward=False)
trainer.stop()
def _import_mbmpo():
import ray.rllib.algorithms.mbmpo as mbmpo
return mbmpo.MBMPO, mbmpo.MBMPOConfig().to_dict()