def good_inverted_car(directory="./Results/Car/Noisy_500/"):
rccar = RCCarBarriers(noise=0.)
policies = [
PolicyLoader("models/noisy/" + path) for path in ['good', 'other']
]
domain = RLPyEnv(rccar)
env = HRLEnv(domain, policies)
policy = CategoricalMLPPolicy(env_spec=env.spec, )
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=0.9,
step_size=0.001,
# plot=True,
)
assert False, "Make sure to change logging directory before rerunning this experiment"
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
script="scripts/run_experiment_lite_rl.py",
log_dir=directory,
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
seed=1,
# plot=True,
)
def test(num=1, path="./Results/Tmp", save=False):
policies = [
PolicyLoader("models/Acrobot/" + path)
for path in ['Mass2_Light', 'Mass1_Light', 'Mass1_Heavy'][:]
]
# directory = os.path.join(directory, exp_name)
acrobot = ModifiedAcrobot()
acrobot.dt = 0.1
acrobot.episodeCap = 1000
# acrobot.torque_noise_max = 0.05
domain = RLPyEnv(acrobot)
env = HRLEnv(domain, policies)
# env = DoublePendulumEnv()
policy = CategoricalMLPPolicy(env_spec=env.spec, hidden_sizes=(8, 8))
# rollout(env, policy)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=env.horizon,
n_itr=50,
discount=0.995,
step_size=0.001,
# plot=True,
)
algo.train()
def noisy_bandits():
rccar = RCCarBarriers(noise=0.1)
policies = [PolicyLoader("models/noisy/" +path) for path in ['good','untrained', 'untrained', 'untrained'] ]
domain = RLPyEnv(rccar)
band = Bandits(policies, domain, N=100, rmax=rccar.GOAL_REWARD, rmin=-rccar.episodeCap - 20)
chosen = band.run()
import joblib; joblib.dump(band.choices, "Results/Bandits/Noisy")
return band
def two_policy_bandits():
rccar = RCCarSlideTurn(noise=0.1) # remove process noise
domain = RLPyEnv(rccar)
policies = [PolicyLoader("models/slideturn_experiment/" + path) for path in ['agent0','agent1'] ]
band = Bandits(policies, domain, N=100, rmax=rccar.GOAL_REWARD, rmin=-rccar.episodeCap)
chosen = band.run()
import joblib; joblib.dump(band.choices, "Results/Bandits/TwoPolicy")
return band
def good_double_bad_car(num=1,
directory="./Results/Car/CarNoisy/",
exp_name="Noisy_2_untrained",
save=False):
rccar = RCCarBarriers(noise=0.1)
policies = [
PolicyLoader("models/noisy/" + path)
for path in ['good', 'untrained', 'untrained']
]
domain = RLPyEnv(rccar)
env = HRLEnv(domain, policies)
# policy = CategoricalMLPPolicy(
# env_spec=env.spec,
# )
# 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=0.9,
# step_size=0.001,
# # plot=True,
# )
policy = CategoricalMLPPolicy(env_spec=env.spec, hidden_sizes=(64, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
for i in range(num):
now = datetime.datetime.now()
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S')
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=env.horizon,
n_itr=600,
discount=0.9,
step_size=0.0001,
# plot=True,
)
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
script="scripts/run_experiment_lite_rl.py",
exp_name=exp_name + timestamp,
log_dir=os.path.join(directory, exp_name +
timestamp) if save else "./Results/Tmp",
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
# plot=True,
)
def good_x_cars(num_agents=5,
directory="./Final_Results/Car/CarNoisyAgents/",
exp_name="NoisyTest",
save=False):
rccar = RCCarBarriers(noise=0.1)
policies = [
PolicyLoader("models/noisy/" + path) for path in [
'good', 'untrained', 'untrained', 'untrained', 'untrained',
'untrained'
]
][:1 + num_agents]
domain = RLPyEnv(rccar)
env = HRLEnv(domain, policies)
policy = CategoricalMLPPolicy(env_spec=env.spec, hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
exp_name = exp_name + str(num_agents)
directory = os.path.join(directory, exp_name)
for i in range(3):
now = datetime.datetime.now()
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S')
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=2000,
max_path_length=env.horizon,
n_itr=500,
discount=.995,
step_size=0.001,
# plot=True,
)
# algo.train()
# rollout(env, policy)
try:
os.mkdir(directory)
except Exception:
pass
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
script="scripts/run_experiment_lite_rl.py",
exp_name=exp_name,
log_dir=os.path.join(directory, timestamp)
if save else "./Results/Tmp",
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
# plot=True,
)
def slideturn_noisy(val=0.1,
directory="./Results/Car/NoisyObs500/",
exp_name="Cap_",
save=False):
policies = [
PolicyLoader("models/slideturn_experiment/" + path)
for path in ['agent0', 'agent1']
]
rccar = RCCarSlideTurn(noise=0.) # remove process noise
domain = RLPyEnv(rccar)
original_env = HRLEnv(domain, policies)
env = NoisyObservationEnv(original_env, obs_noise=val)
policy = CategoricalMLPPolicy(env_spec=env.spec, )
baseline = LinearFeatureBaseline(env_spec=env.spec)
dir_name = os.path.join(directory, exp_name)
for i in range(1):
now = datetime.datetime.now()
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S')
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=env.horizon,
n_itr=500,
discount=0.9,
step_size=0.01,
# plot=True,
)
# algo.train()
# rollout(env, policy)
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
script="scripts/run_experiment_lite_rl.py",
exp_name=exp_name + timestamp,
log_dir=os.path.join(dir_name, timestamp)
if save else './Results/Tmp2',
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
# plot=True,
)
def slideturn_turn_only(num=1, directory="./Final_Results/Car/SlideTurn/", exp_name="Turn", save=True):
policies = [PolicyLoader("models/slideturn_experiment/" + path) for path in ['agent1'] ]
directory = os.path.join(directory, exp_name)
for i in range(num):
rccar = RCCarSlideTurn(noise=0.1)
now = datetime.datetime.now()
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S')
domain = RLPyEnv(rccar)
env = HRLEnv(domain, policies)
policy = CategoricalMLPPolicy(
env_spec=env.spec,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=1000,
max_path_length=env.horizon,
n_itr=500 + 3,
discount=0.995,
step_size=0.001,
# plot=True,
)
# algo.train()
# rollout(env, policy)
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
script="scripts/run_experiment_lite_rl.py",
exp_name=exp_name,
log_dir=os.path.join(directory, timestamp) if save else './Results/Tmp',
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
# plot=True,
)