def test_ceppo(local_mode=False):
_base(
CEPPOTrainer,
local_mode,
extra_config={
"mode": tune.grid_search(
[
# DISABLE,
DISABLE_AND_EXPAND,
REPLAY_VALUES,
# NO_REPLAY_VALUES,
# DIVERSITY_ENCOURAGING,
# DIVERSITY_ENCOURAGING_NO_RV,
# DIVERSITY_ENCOURAGING_DISABLE,
# DIVERSITY_ENCOURAGING_DISABLE_AND_EXPAND, CURIOSITY,
# CURIOSITY_NO_RV,
# CURIOSITY_DISABLE,
# CURIOSITY_DISABLE_AND_EXPAND,
# CURIOSITY_KL,
# CURIOSITY_KL_NO_RV,
# CURIOSITY_KL_DISABLE,
# CURIOSITY_KL_DISABLE_AND_EXPAND
]
),
"num_cpus_per_worker": 0.5,
"num_workers": 1,
# new config:
"clip_action_prob_kl": 0.0
},
# config={"mode": DIVERSITY_ENCOURAGING},
env_name="Pendulum-v0",
t=10000
)
def test_dece(config={}, local_mode=False, t=2000, **kwargs):
_base(
trainer=DECETrainer,
local_mode=local_mode,
extra_config=config,
env_name="Pendulum-v0",
t=t,
**kwargs
)
def test_vtrace_single_agent(local_mode=False):
_base(
trainer=DECETrainer,
local_mode=local_mode,
extra_config={
REPLAY_VALUES: tune.grid_search([True, False]),
'sample_batch_size': 50,
'train_batch_size': 200,
'num_sgd_iter': 10,
'sgd_minibatch_size': 50
},
env_name=FourWayGridWorld,
t=20000,
num_agents=1
)
def mock_experiment(lm=False):
_base(
trainer=DECETrainer,
local_mode=lm,
extra_config={
DELAY_UPDATE: tune.grid_search([True, False]),
REPLAY_VALUES: tune.grid_search([True, False]),
'sample_batch_size': 20,
'sgd_minibatch_size': 100,
'train_batch_size': 500,
},
env_name=FourWayGridWorld,
t={'timesteps_total': 5000},
num_agents=tune.grid_search([1, 5])
)
def test_vtrace(local_mode=False, hard=False):
_base(
trainer=DECETrainer,
local_mode=local_mode,
extra_config={
REPLAY_VALUES: True,
'sample_batch_size': 50 if hard else 8,
'train_batch_size': 450 if hard else 96,
'num_sgd_iter': 10 if hard else 2,
"sgd_minibatch_size": 150 if hard else 3 * 8,
'model': {
'fcnet_hiddens': [16, 16]
},
'seed': 0
# 'lr': 5e-3,
},
env_name=FourWayGridWorld,
t=100000
)
def no_replay_values_batch_size_bug(lm=False):
_base(
trainer=DECETrainer,
local_mode=lm,
extra_config={
REPLAY_VALUES: tune.grid_search([True, False]),
CONSTRAIN_NOVELTY: tune.grid_search(['soft', 'hard', None]),
'num_envs_per_worker': 4,
'sample_batch_size': 20,
'sgd_minibatch_size': 100,
'train_batch_size': 1000,
"num_cpus_per_worker": 1,
"num_cpus_for_driver": 1,
'num_workers': 2,
},
env_name=FourWayGridWorld,
t=1000000,
num_agents=tune.grid_search([5])
)
def regression_test(local_mode=False):
_base(
trainer=DECETrainer,
local_mode=local_mode,
extra_config={
REPLAY_VALUES: tune.grid_search([True, False]),
# "normalize_advantage": tune.grid_search([True, False]),
# 'use_vtrace': tune.grid_search([True]),
'sample_batch_size': 128,
'train_batch_size': 512,
'sgd_minibatch_size': 128,
'num_sgd_iter': 10,
USE_BISECTOR: False,
'seed': tune.grid_search([432, 1920]),
# 'lr': 5e-3,
},
# env_name="Pendulum-v0",
# env_name="CartPole-v0",
env_name=FourWayGridWorld,
t={'time_total_s': 300},
# t={'timesteps_total': 300000},
num_agents=1
)
def test_single_agent(local_mode=False):
_base(CEPPOTrainer, local_mode, dict(mode=DISABLE), num_agents=1)