def create_mlp_deterministic_policy(self, env_spec, name='mlp_policy'):
policy = DeterministicMLPPolicy(env_spec=env_spec,
name=name,
name_scope=name,
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS":
env_spec.flat_action_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}],
output_high=None,
output_low=None,
output_norm=None,
input_norm=None,
reuse=False)
return policy, locals()
def create_ddpg(self, env_id='Pendulum-v0', name='ddpg'):
env = make(env_id)
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + 'mlp_q',
name=name + 'mlp_q',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
self.assertTrue(len(mlp_q.parameters('tf_var_list')) == 4)
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + 'mlp_policy',
name=name + 'mlp_policy',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS":
env_spec.flat_action_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}],
reuse=False)
self.assertTrue(len(policy.parameters('tf_var_list')) == 4)
ddpg = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": 10000,
"GAMMA": 0.999,
"CRITIC_LEARNING_RATE": 0.001,
"ACTOR_LEARNING_RATE": 0.001,
"DECAY": 0.5,
"BATCH_SIZE": 50,
"TRAIN_ITERATION": 1,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=policy,
name=name,
replay_buffer=None)
return ddpg, locals()
def task_fn():
env = make('Pendulum-v0')
name = 'mb_test'
env_spec = env.env_spec
model_path = '/home/yitongx/Documents/baconian-project/experiments/log'
cyber = PendulumnCyber(env=env, epoch_to_use=60, use_traj_input=False, use_mbmf=True, \
model_path=model_path)
mlp_config = [{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 32,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "2",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "3",
"N_UNITS": 8,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "TANH",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}]
mlp_q = MLPQValueFunction(env_spec=env_spec,
name=name + '_mlp_q',
name_scope=name + '_mlp_q',
output_high=env.action_space.high,
mlp_config=mlp_config)
mlp_policy = DeterministicMLPPolicy(env_spec=env_spec,
name=name + '_mlp_policy',
name_scope=name + '_mlp_policy',
output_high=env.observation_space.high,
mlp_config=mlp_config,
reuse=False)
polyak = 0.995
gamma = 0.99
noise_scale = 0.5
noise_decay = 0.999 # default 0.995
batch_size = 128
actor_lr = 0.001 # default 0.001
critic_lr = 0.001 # default 0.001
buffer_size = 100000
total_steps = 500000 # default 1000000
max_step_per_episode = 500 # reset env when counter > max_step_per_episode
train_after_step = 10000 # default 10000
train_every_step = 1
train_iter_per_call = 1
test_after_step = 10000
test_every_step = 1000
num_test = 10
algo = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": buffer_size,
"GAMMA": gamma,
"CRITIC_LEARNING_RATE": critic_lr,
"ACTOR_LEARNING_RATE": actor_lr,
"DECAY": polyak,
"BATCH_SIZE": batch_size,
"TRAIN_ITERATION": train_iter_per_call,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=mlp_policy,
name=name + '_ddpg',
replay_buffer=None)
step_counter = SinglentonStepCounter(-1)
noise_adder = AgentActionNoiseWrapper(
noise=UniformNoise(scale=noise_scale),
action_weight_scheduler=ConstantScheduler(1.),
noise_weight_scheduler=DDPGNoiseScheduler(
train_every_step=train_every_step,
train_after_step=train_after_step,
noise_decay=noise_decay,
step_counter=step_counter))
agent = DDPG_Agent(env=env,
algo=algo,
env_spec=env_spec,
noise_adder=noise_adder,
name=name + '_agent')
flow = create_train_test_flow(env=env,
cyber=cyber,
agent=agent,
num_test=num_test,
total_steps=total_steps,
max_step_per_episode=max_step_per_episode,
train_after_step=train_after_step,
test_after_step=test_after_step,
train_every_step=train_every_step,
test_every_step=test_every_step,
train_func_and_args=(agent.train, (),
dict()),
test_func_and_args=(agent.test, (), dict()),
sample_func_and_args=(agent.sample, (),
dict()),
flow_type='DDPG_TrainTestFlow')
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
def task_fn():
env = make('Pendulum-v0')
name = 'demo_exp'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": env_spec.flat_action_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}],
reuse=False)
ddpg = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": 10000,
"GAMMA": 0.999,
"CRITIC_LEARNING_RATE": 0.001,
"ACTOR_LEARNING_RATE": 0.001,
"DECAY": 0.5,
"BATCH_SIZE": 50,
"TRAIN_ITERATION": 1,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=policy,
name=name + '_ddpg',
replay_buffer=None)
agent = Agent(
env=env,
env_spec=env_spec,
algo=ddpg,
algo_saving_scheduler=PeriodicalEventSchedule(
t_fn=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
trigger_every_step=20,
after_t=10),
name=name + '_agent',
exploration_strategy=EpsilonGreedy(action_space=env_spec.action_space,
init_random_prob=0.5))
flow = create_train_test_flow(
test_every_sample_count=10,
train_every_sample_count=10,
start_test_after_sample_count=5,
start_train_after_sample_count=5,
train_func_and_args=(agent.train, (), dict()),
test_func_and_args=(agent.test, (), dict(sample_count=10)),
sample_func_and_args=(agent.sample, (),
dict(sample_count=100,
env=agent.env,
in_which_status='TRAIN',
store_flag=True)))
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
def task_fn():
env = make('Pendulum-v0')
name = 'demo_exp'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": env_spec.flat_action_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}],
reuse=False)
ddpg = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": 10000,
"GAMMA": 0.999,
"CRITIC_LEARNING_RATE": 0.001,
"ACTOR_LEARNING_RATE": 0.001,
"DECAY": 0.5,
"BATCH_SIZE": 50,
"TRAIN_ITERATION": 1,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=policy,
name=name + '_ddpg',
replay_buffer=None)
mlp_dyna_list = []
for i in range(10):
mlp_dyna = ContinuousMLPGlobalDynamicsModel(
env_spec=env_spec,
name_scope=name + '_mlp_dyna_{}'.format(i),
name=name + '_mlp_dyna_{}'.format(i),
learning_rate=0.01,
state_input_scaler=RunningStandardScaler(
dims=env_spec.flat_obs_dim),
action_input_scaler=RunningStandardScaler(
dims=env_spec.flat_action_dim),
output_delta_state_scaler=RunningStandardScaler(
dims=env_spec.flat_obs_dim),
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"L1_NORM": 0.0,
"L2_NORM": 0.0,
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"L1_NORM": 0.0,
"L2_NORM": 0.0,
"N_UNITS": env_spec.flat_obs_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
mlp_dyna_list.append(mlp_dyna)
dyna_ensemble_model = ModelEnsemble(n_models=10,
model=mlp_dyna_list,
prediction_type='random',
env_spec=env_spec)
algo = ModelEnsembleAlgo(env_spec=env_spec,
model_free_algo=ddpg,
dynamics_model=dyna_ensemble_model,
config_or_config_dict=dict(
dynamics_model_train_iter=10,
model_free_algo_train_iter=10,
validation_trajectory_count=2,
))
# For examples only, we use random reward function and terminal function with fixed episode length.
algo.set_terminal_reward_function_for_dynamics_env(
terminal_func=FixedEpisodeLengthTerminalFunc(
max_step_length=env.unwrapped._max_episode_steps,
step_count_fn=algo.dynamics_env.total_step_count_fn),
reward_func=PendulumRewardFunc())
agent = Agent(
env=env,
env_spec=env_spec,
algo=algo,
algo_saving_scheduler=PeriodicalEventSchedule(
t_fn=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
trigger_every_step=200,
after_t=10),
name=name + '_agent',
exploration_strategy=EpsilonGreedy(action_space=env_spec.action_space,
init_random_prob=0.5))
# we can easily reuse the dyna training flow to implement the Model-ensemble training flow.
flow = create_dyna_flow(
train_algo_func=(agent.train, (), dict(state='state_agent_training')),
train_algo_from_synthesized_data_func=(
agent.train, (), dict(state='state_agent_training')),
train_dynamics_func=(agent.train, (),
dict(state='state_dynamics_training')),
test_algo_func=(agent.test, (), dict(sample_count=10)),
test_dynamics_func=(agent.algo.test_dynamics, (),
dict(sample_count=10, env=env)),
sample_from_real_env_func=(agent.sample, (),
dict(sample_count=10,
env=agent.env,
store_flag=True)),
sample_from_dynamics_env_func=(agent.sample, (),
dict(sample_count=10,
env=agent.algo.dynamics_env,
store_flag=True)),
# set this to large enough so agent only use data from dynamics env.
train_algo_every_real_sample_count_by_data_from_real_env=100,
train_algo_every_real_sample_count_by_data_from_dynamics_env=100,
test_algo_every_real_sample_count=100,
test_dynamics_every_real_sample_count=100,
train_dynamics_ever_real_sample_count=100,
start_train_algo_after_sample_count=1,
start_train_dynamics_after_sample_count=1,
start_test_algo_after_sample_count=1,
start_test_dynamics_after_sample_count=1,
warm_up_dynamics_samples=100)
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name + '_exp')
experiment.run()
def hopper_task_fn():
exp_config = HOPPER_BENCHMARK_CONFIG_DICT
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT',
exp_config['DEFAULT_EXPERIMENT_END_POINT'])
env = make('Hopper-v2')
name = 'benchmark'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
**exp_config['MLP_V'])
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
output_low=env_spec.action_space.low,
output_high=env_spec.action_space.high,
**exp_config['POLICY'],
reuse=False)
ddpg = DDPG(env_spec=env_spec,
policy=policy,
value_func=mlp_q,
name=name + '_ddpg',
**exp_config['DDPG'])
agent = Agent(env=env,
env_spec=env_spec,
algo=ddpg,
exploration_strategy=None,
noise_adder=AgentActionNoiseWrapper(
noise=OUNoise(theta=0.15, sigma=0.3),
noise_weight_scheduler=ConstantScheduler(1),
action_weight_scheduler=ConstantScheduler(1),
),
name=name + '_agent')
flow = TrainTestFlow(
train_sample_count_func=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
config_or_config_dict=exp_config['TrainTestFlow']
['config_or_config_dict'],
func_dict={
'test': {
'func':
agent.test,
'args':
list(),
'kwargs':
dict(sample_count=exp_config['TrainTestFlow']
['TEST_SAMPLES_COUNT']),
},
'train': {
'func': agent.train,
'args': list(),
'kwargs': dict(),
},
'sample': {
'func':
agent.sample,
'args':
list(),
'kwargs':
dict(sample_count=exp_config['TrainTestFlow']
['TRAIN_SAMPLES_COUNT'],
env=agent.env,
in_which_status='TRAIN',
store_flag=True),
},
})
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
def mountiancar_task_fn():
exp_config = MOUNTAIN_CAR_CONTINUOUS_BENCHMARK_CONFIG_DICT
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT',
exp_config['DEFAULT_EXPERIMENT_END_POINT'])
env = make('MountainCarContinuous-v0')
name = 'benchmark'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
**exp_config['MLPQValueFunction'])
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
output_low=env_spec.action_space.low,
output_high=env_spec.action_space.high,
**exp_config['DeterministicMLPPolicy'],
reuse=False)
ddpg = DDPG(env_spec=env_spec,
policy=policy,
value_func=mlp_q,
name=name + '_ddpg',
**exp_config['DDPG'])
n_actions = env.action_space.shape[0]
agent = Agent(env=env,
env_spec=env_spec,
algo=ddpg,
exploration_strategy=None,
noise_adder=AgentActionNoiseWrapper(
noise=OrnsteinUhlenbeckActionNoise(
mu=np.zeros(n_actions),
sigma=0.5 * np.ones(n_actions)),
noise_weight_scheduler=ConstantScheduler(value=1),
action_weight_scheduler=ConstantScheduler(value=1.0)),
reset_noise_every_terminal_state=True,
name=name + '_agent')
flow = TrainTestFlow(
train_sample_count_func=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
config_or_config_dict=exp_config['TrainTestFlow']
['config_or_config_dict'],
func_dict={
'test': {
'func':
agent.test,
'args':
list(),
'kwargs':
dict(sample_count=exp_config['TrainTestFlow']
['TEST_SAMPLES_COUNT']),
},
'train': {
'func': agent.train,
'args': list(),
'kwargs': dict(),
},
'sample': {
'func':
agent.sample,
'args':
list(),
'kwargs':
dict(sample_count=exp_config['TrainTestFlow']
['TRAIN_SAMPLES_COUNT'],
env=agent.env,
in_which_status='TRAIN',
store_flag=True),
},
})
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
def pendulum_task_fn():
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT',
exp_config['DEFAULT_EXPERIMENT_END_POINT'])
env = make('Pendulum-v0')
name = 'benchmark'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
**exp_config['MLPQValueFunction'])
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
output_low=env_spec.action_space.low,
output_high=env_spec.action_space.high,
**exp_config['DeterministicMLPPolicy'],
reuse=False)
ddpg = DDPG(env_spec=env_spec,
policy=policy,
value_func=mlp_q,
name=name + '_ddpg',
**exp_config['DDPG'])
mlp_dyna = ContinuousMLPGlobalDynamicsModel(env_spec=env_spec,
name_scope=name + '_mlp_dyna',
name=name + '_mlp_dyna',
**exp_config['DynamicsModel'])
algo = Dyna(env_spec=env_spec,
name=name + '_dyna_algo',
model_free_algo=ddpg,
dynamics_model=mlp_dyna,
config_or_config_dict=dict(dynamics_model_train_iter=10,
model_free_algo_train_iter=10))
algo.set_terminal_reward_function_for_dynamics_env(
terminal_func=FixedEpisodeLengthTerminalFunc(
max_step_length=env.unwrapped._max_episode_steps,
step_count_fn=algo.dynamics_env.total_step_count_fn),
reward_func=REWARD_FUNC_DICT['Pendulum-v0']())
agent = Agent(env=env,
env_spec=env_spec,
algo=algo,
exploration_strategy=None,
noise_adder=AgentActionNoiseWrapper(
noise=NormalActionNoise(),
noise_weight_scheduler=ConstantSchedule(value=0.3),
action_weight_scheduler=ConstantSchedule(value=1.0)),
name=name + '_agent')
flow = DynaFlow(
train_sample_count_func=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
config_or_config_dict=exp_config['DynaFlow'],
func_dict={
'train_algo': {
'func': agent.train,
'args': list(),
'kwargs': dict(state='state_agent_training')
},
'train_algo_from_synthesized_data': {
'func': agent.train,
'args': list(),
'kwargs': dict(state='state_agent_training', train_iter=1)
},
'train_dynamics': {
'func': agent.train,
'args': list(),
'kwargs': dict(state='state_dynamics_training')
},
'test_algo': {
'func': agent.test,
'args': list(),
'kwargs': dict(sample_count=1, sample_trajectory_flag=True)
},
'test_dynamics': {
'func': agent.algo.test_dynamics,
'args': list(),
'kwargs': dict(sample_count=10, env=env)
},
'sample_from_real_env': {
'func':
agent.sample,
'args':
list(),
'kwargs':
dict(sample_count=10,
env=agent.env,
in_which_status='TRAIN',
store_flag=True)
},
'sample_from_dynamics_env': {
'func':
agent.sample,
'args':
list(),
'kwargs':
dict(sample_count=50,
sample_type='transition',
env=agent.algo.dynamics_env,
in_which_status='TRAIN',
store_flag=False)
}
})
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
def task_fn():
# create the gym environment by make function
env = make('Pendulum-v0')
# give your experiment a name which is used to generate the log path etc.
name = 'demo_exp'
# construct the environment specification
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
# construct the neural network to approximate q function of DDPG
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
# construct the neural network to approximate policy for DDPG
policy = DeterministicMLPPolicy(env_spec=env_spec,
name_scope=name + '_mlp_policy',
name=name + '_mlp_policy',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": env_spec.flat_action_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}],
reuse=False)
# construct the DDPG algorithms
ddpg = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": 10000,
"GAMMA": 0.999,
"CRITIC_LEARNING_RATE": 0.001,
"ACTOR_LEARNING_RATE": 0.001,
"DECAY": 0.5,
"BATCH_SIZE": 50,
"TRAIN_ITERATION": 1,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=policy,
name=name + '_ddpg',
replay_buffer=None)
# construct a neural network based global dynamics model to approximate the state transition of environment
mlp_dyna = ContinuousMLPGlobalDynamicsModel(
env_spec=env_spec,
name_scope=name + '_mlp_dyna',
name=name + '_mlp_dyna',
learning_rate=0.01,
state_input_scaler=RunningStandardScaler(dims=env_spec.flat_obs_dim),
action_input_scaler=RunningStandardScaler(
dims=env_spec.flat_action_dim),
output_delta_state_scaler=RunningStandardScaler(
dims=env_spec.flat_obs_dim),
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"L1_NORM": 0.0,
"L2_NORM": 0.0,
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"L1_NORM": 0.0,
"L2_NORM": 0.0,
"N_UNITS": env_spec.flat_obs_dim,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
# finally, construct the Dyna algorithms with a model free algorithm DDGP, and a NN model.
algo = Dyna(env_spec=env_spec,
name=name + '_dyna_algo',
model_free_algo=ddpg,
dynamics_model=mlp_dyna,
config_or_config_dict=dict(dynamics_model_train_iter=10,
model_free_algo_train_iter=10))
# To make the NN based dynamics model a proper environment so be a sampling source for DDPG, reward function and
# terminal function need to be set.
# For examples only, we use random reward function and terminal function with fixed episode length.
algo.set_terminal_reward_function_for_dynamics_env(
terminal_func=FixedEpisodeLengthTerminalFunc(
max_step_length=env.unwrapped._max_episode_steps,
step_count_fn=algo.dynamics_env.total_step_count_fn),
reward_func=RandomRewardFunc())
# construct agent with additional exploration strategy if needed.
agent = Agent(
env=env,
env_spec=env_spec,
algo=algo,
algo_saving_scheduler=PeriodicalEventSchedule(
t_fn=lambda: get_global_status_collect()
('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
trigger_every_step=20,
after_t=10),
name=name + '_agent',
exploration_strategy=EpsilonGreedy(action_space=env_spec.action_space,
init_random_prob=0.5))
# construct the training flow, called Dyna flow. It defines how the training proceed, and the terminal condition
flow = create_dyna_flow(
train_algo_func=(agent.train, (), dict(state='state_agent_training')),
train_algo_from_synthesized_data_func=(
agent.train, (), dict(state='state_agent_training')),
train_dynamics_func=(agent.train, (),
dict(state='state_dynamics_training')),
test_algo_func=(agent.test, (), dict(sample_count=1)),
test_dynamics_func=(agent.algo.test_dynamics, (),
dict(sample_count=10, env=env)),
sample_from_real_env_func=(agent.sample, (),
dict(sample_count=10,
env=agent.env,
store_flag=True)),
sample_from_dynamics_env_func=(agent.sample, (),
dict(sample_count=10,
env=agent.algo.dynamics_env,
store_flag=True)),
train_algo_every_real_sample_count_by_data_from_real_env=40,
train_algo_every_real_sample_count_by_data_from_dynamics_env=40,
test_algo_every_real_sample_count=40,
test_dynamics_every_real_sample_count=40,
train_dynamics_ever_real_sample_count=20,
start_train_algo_after_sample_count=1,
start_train_dynamics_after_sample_count=1,
start_test_algo_after_sample_count=1,
start_test_dynamics_after_sample_count=1,
warm_up_dynamics_samples=1)
# construct the experiment
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name + '_exp')
# run!
experiment.run()
def test_DDPG_1(self):
self.LogSetup()
env = make('Pendulum-v0')
name = 'mb_test'
env_spec = env.env_spec
cyber = PendulumnCyber(env=env, epoch_to_use=60, use_traj_input=False, use_mbmf=True, \
model_path='/home/yitongx/Documents/baconian-project/experiments/log')
actor_policy_mlp_config = [{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 32,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "2",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "3",
"N_UNITS": 8,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "TANH",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}]
mlp_q = MLPQValueFunction(env_spec=env_spec,
name=name + '_mlp_q',
name_scope=name + '_mlp_q',
output_high=env.action_space.high,
mlp_config=actor_policy_mlp_config)
mlp_policy = DeterministicMLPPolicy(
env_spec=env_spec,
name=name + '_mlp_policy',
name_scope=name + '_mlp_policy',
output_high=env.observation_space.high,
mlp_config=actor_policy_mlp_config,
reuse=False)
polyak = 0.995
gamma = 0.99
batch_size = 128
actor_lr = 0.001
critic_lr = 0.001
train_iter_per_call = 1
buffer_size = 100000
total_steps = 100000 # default 1000000
max_step_per_episode = 500 # reset env when counter > max_step_per_episode
train_after_step = 10000
test_after_step = 10000
train_every_step = 1
test_every_step = 1000
num_test = 10
algo = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": buffer_size,
"GAMMA": gamma,
"CRITIC_LEARNING_RATE": critic_lr,
"ACTOR_LEARNING_RATE": actor_lr,
"DECAY": polyak,
"BATCH_SIZE": batch_size,
"TRAIN_ITERATION": train_iter_per_call,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=mlp_policy,
name=name + '_ddpg',
replay_buffer=None)
algo.init()
buffer = TransitionData(env_spec=env_spec,
obs_shape=env_spec.obs_shape,
action_shape=env_spec.action_shape)
for i in range(100): # num_trajectory
obs = env.reset()
for j in range(1000):
action = env.action_space.sample()
obs_, reward, done, info = env.step(action)
buffer.append(obs, action, obs_, done, reward)
if done:
break
else:
obs = obs_
algo.append_to_memory(buffer)
for i in range(10):
res = algo.train()
print(res)
obs = env.reset()
act = env.action_space.sample()
obs_, reward, done, info = cyber.step(obs, act)
print('obs_', obs_)
obs = env.observation_space.sample()
obs_batch = np.array([])
for i in range(5):
obs_batch = np.concatenate(
(obs_batch, env.observation_space.sample()), axis=0)
# print(obs_batch.shape)
act_batch = algo.predict(obs_batch)
print(act_batch.shape)
print(act_batch)
print('====> Test')
act = algo.predict(obs)
print(act.shape)
obs = env.reset()
print(obs.shape)
obs_, reward, done, info = cyber.step(obs, act)
def test_DDPG_2(self):
self.LogSetup()
env = make('Pendulum-v0')
name = 'mb_test'
env_spec = env.env_spec
cyber = PendulumnCyber(env=env, epoch_to_use=60, use_traj_input=False, use_mbmf=True, \
model_path='/home/yitongx/Documents/baconian-project/experiments/log')
actor_policy_mlp_config = [{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 32,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "2",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "3",
"N_UNITS": 8,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "TANH",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}]
mlp_q = MLPQValueFunction(env_spec=env_spec,
name=name + '_mlp_q',
name_scope=name + '_mlp_q',
output_high=env.action_space.high,
mlp_config=actor_policy_mlp_config)
mlp_policy = DeterministicMLPPolicy(
env_spec=env_spec,
name=name + '_mlp_policy',
name_scope=name + '_mlp_policy',
output_high=env.observation_space.high,
mlp_config=actor_policy_mlp_config,
reuse=False)
polyak = 0.995
gamma = 0.99
noise_scale = 0.5
noise_decay = 0.995
batch_size = 128
actor_lr = 0.001
critic_lr = 0.001
train_iter_per_call = 1
buffer_size = 100000
total_steps = 100000 # default 1000000
max_step_per_episode = 500 # reset env when counter > max_step_per_episode
train_after_step = 10000 # default 10000
train_every_step = 1
test_after_step = 10000
test_every_step = 1000
num_test = 10
algo = DDPG(env_spec=env_spec,
config_or_config_dict={
"REPLAY_BUFFER_SIZE": buffer_size,
"GAMMA": gamma,
"CRITIC_LEARNING_RATE": critic_lr,
"ACTOR_LEARNING_RATE": actor_lr,
"DECAY": polyak,
"BATCH_SIZE": batch_size,
"TRAIN_ITERATION": train_iter_per_call,
"critic_clip_norm": 0.1,
"actor_clip_norm": 0.1,
},
value_func=mlp_q,
policy=mlp_policy,
name=name + '_ddpg',
replay_buffer=None)
step_counter = SinglentonStepCounter(-1)
noise_adder = AgentActionNoiseWrapper(
noise=UniformNoise(scale=noise_scale),
action_weight_scheduler=ConstantScheduler(1.),
noise_weight_scheduler=DDPGNoiseScheduler(
train_every_step=train_every_step,
train_after_step=train_after_step,
noise_decay=noise_decay,
step_counter=step_counter))
agent = DDPG_Agent(env=env,
algo=algo,
env_spec=env_spec,
noise_adder=noise_adder,
name=name + '_agent')
agent.init()
test_reward = []
data_sample = []
obs, ep_ret, ep_len = env.reset(), 0, 0
for step in range(total_steps):
step_counter.increase(1)
act = agent.predict(obs=obs)
obs_, reward, done, _ = cyber.step(obs, act)
_buffer = TransitionData(env_spec=env_spec,
obs_shape=env_spec.obs_shape,
action_shape=env_spec.action_shape)
_buffer.append(obs, act, obs_, done, reward)
agent.algo.append_to_memory(_buffer)
ep_ret += reward
ep_len += 1
if done or ep_len > max_step_per_episode:
obs, ep_ret, ep_len = env.reset(), 0, 0
else:
obs = obs_
if step > train_after_step and step % train_every_step == 0:
agent.train()
if step > test_after_step and step % test_every_step == 0:
data_sample, test_reward = agent.test(
env=env,
cyber=cyber,
data_sample=data_sample,
test_reward=test_reward,
num_test=num_test,
max_step_per_episode=max_step_per_episode)