Closed haanvid closed 4 years ago
I used the code below to run the TRPO on the mountain-car environment.
from rllab.algos.trpo import TRPO
from rllab.baselines.linear_feature_baseline import LinearFeatureBaseline
from rllab.envs.gym_env import GymEnv
from rllab.envs.normalized_env import normalize
from rllab.misc.instrument import run_experiment_lite
from rllab.policies.categorical_mlp_policy import CategoricalMLPPolicy
def run_task(*_):
# Please note that different environments with different action spaces may
# require different policies. For example with a Discrete action space, a
# CategoricalMLPPolicy works, but for a Box action space may need to use
# a GaussianMLPPolicy (see the trpo_gym_pendulum.py example)
env = normalize(GymEnv("MountainCar-v0"))
policy = CategoricalMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(32, 32)
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=env.horizon,
n_itr=500,
discount=0.99,
step_size=0.01,
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train()
run_experiment_lite(
run_task,
# Number of parallel workers for sampling
n_parallel=1,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
seed=1,
# plot=True,
)
I've uploaded the rendered result here:
Does the OpenAI version use the same policy network, baselines, and hyperparameters as your rllab implementation?
Hi, Haanvid. I met the same problem. Have you solved the problem yet? @haanvid
Hi all -- this repository is unmaintained, but the spirit of rllab lives on in the garage project at https://github.com/rlworkgroup/garage.
@HuangJiaLian After experiencing some issues, I've switched to using stable-baselines. @ryanjulian I guess they are a bit different. But given the fact that TRPO is one of the stable RL algorithms, it should work on toy domains such as mountain-car.
when the rllab trpo code is applied to the mountain-car env., it does not climb the mountain well until 500 iterations.
It is strange since the TRPO algorithm implemented by OpenAI (https://github.com/openai/baselines/tree/master/baselines/trpo_mpi) climbs the mountain well.