def test_eps_greedy(self):
dqn, locals = self.create_dqn()
dqn.init()
env = locals['env']
eps = EpsilonGreedy(action_space=dqn.env_spec.action_space,
init_random_prob=0.5)
st = env.reset()
for i in range(100):
ac = eps.predict(obs=st, sess=self.sess, batch_flag=False, algo=dqn)
st_new, re, done, _ = env.step(action=ac)
self.assertTrue(env.action_space.contains(ac))
def mountaincar_task_fn():
exp_config = MOUNTAINCAR_BENCHMARK_CONFIG_DICT
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT',
exp_config['DEFAULT_EXPERIMENT_END_POINT'])
env = make('MountainCar-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'])
dqn = DQN(env_spec=env_spec,
name=name + '_dqn',
value_func=mlp_q,
**exp_config['DQN'])
agent = Agent(env=env, env_spec=env_spec,
algo=dqn,
name=name + '_agent',
exploration_strategy=EpsilonGreedy(action_space=env_spec.action_space,
prob_scheduler=LinearScheduler(
t_fn=lambda: get_global_status_collect()(
'TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
**exp_config['EpsilonGreedy']['LinearScheduler']),
**exp_config['EpsilonGreedy']['config_or_config_dict']))
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 test_eps_with_scheduler(self):
dqn, locals = self.create_dqn()
env = locals['env']
def func():
global x
return x
dqn.init()
eps = EpsilonGreedy(action_space=dqn.env_spec.action_space,
prob_scheduler=LinearScheduler(initial_p=1.0, t_fn=func, schedule_timesteps=10,
final_p=0.0),
init_random_prob=1.0)
st = env.reset()
for i in range(10):
global x
ac = eps.predict(obs=st, sess=self.sess, batch_flag=False, algo=dqn)
st_new, re, done, _ = env.step(action=ac)
self.assertAlmostEqual(eps.parameters('random_prob_func')(), 1.0 - (1.0 - 0.0) / 10 * x)
x += 1
def create_eps(self, env_spec):
return EpsilonGreedy(action_space=env_spec.action_space,
init_random_prob=0.5), locals()
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('Acrobot-v1')
name = 'example_scheduler_'
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
}
])
dqn = DQN(env_spec=env_spec,
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=1000,
GAMMA=0.99,
BATCH_SIZE=10,
LEARNING_RATE=0.001,
TRAIN_ITERATION=1,
DECAY=0.5),
name=name + '_dqn',
value_func=mlp_q)
agent = Agent(env=env, env_spec=env_spec,
algo=dqn,
name=name + '_agent',
algo_saving_scheduler=PeriodicalEventSchedule(
t_fn=lambda: get_global_status_collect()('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
trigger_every_step=20,
after_t=10),
exploration_strategy=EpsilonGreedy(action_space=env_spec.action_space,
prob_scheduler=PiecewiseScheduler(
t_fn=lambda: get_global_status_collect()(
'TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
endpoints=((10, 0.3), (100, 0.1), (200, 0.0)),
outside_value=0.0
),
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,
store_flag=True))
)
experiment = Experiment(
tuner=None,
env=env,
agent=agent,
flow=flow,
name=name + 'experiment_debug'
)
dqn.parameters.set_scheduler(param_key='LEARNING_RATE',
scheduler=LinearScheduler(
t_fn=experiment.TOTAL_AGENT_TRAIN_SAMPLE_COUNT,
schedule_timesteps=GlobalConfig().DEFAULT_EXPERIMENT_END_POINT[
'TOTAL_AGENT_TRAIN_SAMPLE_COUNT'],
final_p=0.0001,
initial_p=0.01))
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_v = MLPVValueFunc(env_spec=env_spec,
name_scope=name + 'mlp_v',
name=name + 'mlp_v',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"L1_NORM": 0.01,
"L2_NORM": 0.01,
"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 = NormalDistributionMLPPolicy(env_spec=env_spec,
name_scope=name + 'mlp_policy',
name=name + 'mlp_policy',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"L1_NORM": 0.01,
"L2_NORM": 0.01,
"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)
ppo = PPO(env_spec=env_spec,
config_or_config_dict={
"gamma": 0.995,
"lam": 0.98,
"policy_train_iter": 10,
"value_func_train_iter": 10,
"clipping_range": None,
"beta": 1.0,
"eta": 50,
"log_var_init": -1.0,
"kl_target": 0.003,
"policy_lr": 0.01,
"value_func_lr": 0.01,
"value_func_train_batch_size": 10,
"lr_multiplier": 1.0
},
value_func=mlp_v,
stochastic_policy=policy,
name=name + 'ppo')
agent = Agent(
env=env,
env_spec=env_spec,
algo=ppo,
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('Acrobot-v1')
name = 'demo_exp'
env.env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
env_spec = env.env_spec
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp_q',
name=name + '_mlp_q',
mlp_config=[
{
"ACT": "TANH",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 64,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
},
{
"ACT": "TANH",
"B_INIT_VALUE": 0.0,
"NAME": "2",
"N_UNITS": 64,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
},
{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "3",
"N_UNITS": 256,
"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
}
])
dqn = DQN(env_spec=env_spec,
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=50000,
GAMMA=0.99,
BATCH_SIZE=32,
LEARNING_RATE=0.001,
TRAIN_ITERATION=1,
DECAY=0),
name=name + '_dqn',
value_func=mlp_q)
epsilon_greedy = EpsilonGreedy(action_space=env_spec.action_space,
prob_scheduler=LinearScheduler(
t_fn=lambda: get_global_status_collect()('TOTAL_AGENT_TRAIN_SAMPLE_COUNT'),
schedule_timesteps=int(0.1 * 100000),
initial_p=1.0,
final_p=0.02),
init_random_prob=0.1)
agent = Agent(env=env,
env_spec=env_spec,
algo=dqn,
name=name + '_agent',
exploration_strategy=epsilon_greedy,
noise_adder=None)
flow = create_train_test_flow(
test_every_sample_count=1000,
train_every_sample_count=1,
start_test_after_sample_count=0,
start_train_after_sample_count=10000,
sample_func_and_args=(agent.sample, (), dict(sample_count=1, env=agent.env, store_flag=True)),
train_func_and_args=(agent.train, (), dict()),
test_func_and_args=(agent.test, (), dict(sample_count=1)),
)
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 = env.env_spec
mlp_dyna = ContinuousMLPGlobalDynamicsModel(env_spec=env_spec,
name_scope=name + '_mlp_dyna',
name=name + '_mlp_dyna',
learning_rate=0.01,
mlp_config=[{
"ACT":
"TANH",
"B_INIT_VALUE":
0.0,
"NAME":
"1",
"L1_NORM":
0.0,
"L2_NORM":
0.0,
"N_UNITS":
128,
"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
}])
algo = ModelPredictiveControl(
dynamics_model=mlp_dyna,
env_spec=env_spec,
config_or_config_dict=dict(SAMPLED_HORIZON=2,
SAMPLED_PATH_NUM=5,
dynamics_model_train_iter=10),
name=name + '_mpc',
policy=UniformRandomPolicy(env_spec=env_spec, name='uni_policy'))
algo.set_terminal_reward_function_for_dynamics_env(
reward_func=RandomRewardFunc(name='reward_func'),
terminal_func=RandomTerminalFunc(name='random_terminal'),
)
agent = Agent(
env=env,
env_spec=env_spec,
algo=algo,
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,
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 task_fn():
env = make('Acrobot-v1')
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
}
])
dqn = DQN(env_spec=env_spec,
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=1000,
GAMMA=0.99,
BATCH_SIZE=10,
LEARNING_RATE=0.01,
TRAIN_ITERATION=1,
DECAY=0.5),
name=name + '_dqn',
value_func=mlp_q)
agent = Agent(env=env, env_spec=env_spec,
algo=dqn,
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():
# 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()