def run_task(v):
record_video = False
import mujoco_envs.pomdp
main_env = GymEnv('Peg3d-v0', record_video=record_video)
# main_env = MultiagentEnv(GymEnv("Swimmer-v1", record_video=record_video))
# replace raw shadow_envs with wrapped envs
main_env._shadow_envs = [TfEnv(ProxyEnv(env)) for env in main_env.shadow_envs]
# main_env._shadow_envs = [TfEnv(normalize(env)) for env in main_env.shadow_envs]
sub_policies = [AutoMLPPolicy(
# sub_policies = [BottleneckAutoMLPPolicy(
name="sub-policy-%s" % i,
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(32, 32) # 32)
) for i,env in enumerate(main_env.shadow_envs)]
# reduces the initialization, to discourage pre-commiting to an action
# for sp in sub_policies:
# import ipdb; ipdb.set_trace()
# sp.get_params()[-3].set_value(sp.get_params()[-3].get_value()*0.01)
policy = MultiMLPPolicy(
name="policy",
env_spec=[env.spec for env in main_env.shadow_envs],
policies=sub_policies
)
baselines = [LinearFeatureBaseline(env_spec=env.spec) for env in main_env.shadow_envs]
# TODO(cathywu) Start with large batch sizes 100-1000 trajectories
algo = MultiTRPO(
env=main_env,
policy=policy,
baselines=baselines,
batch_size=25000,
whole_paths=True,
max_path_length=250,
n_itr=700,
discount=0.995,
step_size=v["step_size"],
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
# NPO_cls=ConsensusNPO,
NPO_cls=NPO,
sample_processor_cls=MultiSampleProcessor,
n_vectorized_envs=40,
)
algo.train()
Walker2DEnv,
SwimmerEnv,
SimpleHumanoidEnv,
InvertedDoublePendulumEnv,
HopperEnv,
HalfCheetahEnv,
PointGatherEnv,
SwimmerGatherEnv,
AntGatherEnv,
PointMazeEnv,
SwimmerMazeEnv,
AntMazeEnv,
])
envs = [cls() for cls in simple_env_classes]
envs.append(ProxyEnv(envs[0]))
envs.append(IdentificationEnv(CartpoleEnv, {}))
envs.append(NoisyObservationEnv(CartpoleEnv()))
envs.append(DelayedActionEnv(CartpoleEnv()))
envs.append(NormalizedEnv(CartpoleEnv()))
envs.append(GymEnv('CartPole-v0'))
@tools.params(*envs)
def test_env(env):
print("Testing", env.__class__)
ob_space = env.observation_space
act_space = env.action_space
ob = env.reset()
assert ob_space.contains(ob)
a = act_space.sample()