def __init__(self,
s_dim,
visual_sources,
visual_resolution,
a_dim,
is_continuous,
ployak=0.995,
actor_lr=5.0e-4,
critic_lr=1.0e-3,
discrete_tau=1.0,
hidden_units={
'actor_continuous': [32, 32],
'actor_discrete': [32, 32],
'q': [32, 32]
},
**kwargs):
super().__init__(
s_dim=s_dim,
visual_sources=visual_sources,
visual_resolution=visual_resolution,
a_dim=a_dim,
is_continuous=is_continuous,
**kwargs)
self.ployak = ployak
self.discrete_tau = discrete_tau
if self.is_continuous:
_actor_net = lambda: rls.actor_dpg(self.feat_dim, self.a_dim, hidden_units['actor_continuous'])
# self.action_noise = rls.NormalActionNoise(mu=np.zeros(self.a_dim), sigma=1 * np.ones(self.a_dim))
self.action_noise = rls.OrnsteinUhlenbeckActionNoise(mu=np.zeros(self.a_dim), sigma=0.2 * np.exp(-self.episode / 10) * np.ones(self.a_dim))
else:
_actor_net = lambda: rls.actor_discrete(self.feat_dim, self.a_dim, hidden_units['actor_discrete'])
self.gumbel_dist = tfp.distributions.Gumbel(0, 1)
self.actor_net = _actor_net()
self.actor_target_net = _actor_net()
self.actor_tv = self.actor_net.trainable_variables
_q_net = lambda : rls.critic_q_one(self.feat_dim, self.a_dim, hidden_units['q'])
self.q_net = _q_net()
self.q_target_net = _q_net()
self.critic_tv = self.q_net.trainable_variables + self.other_tv
self.update_target_net_weights(
self.actor_target_net.weights + self.q_target_net.weights,
self.actor_net.weights + self.q_net.weights
)
self.actor_lr, self.critic_lr = map(self.init_lr, [actor_lr, critic_lr])
self.optimizer_actor, self.optimizer_critic = map(self.init_optimizer, [self.actor_lr, self.critic_lr])
self.model_recorder(dict(
actor=self.actor_net,
critic=self.q_net,
optimizer_actor=self.optimizer_actor,
optimizer_critic=self.optimizer_critic
))
def __init__(self,
s_dim,
a_dim,
is_continuous,
ployak=0.995,
actor_lr=5.0e-4,
critic_lr=1.0e-3,
n=1,
i=0,
hidden_units={
'actor': [32, 32],
'q': [32, 32]
},
**kwargs):
assert is_continuous, 'matd3 only support continuous action space'
raise Exception('MA系列存在问题，还未修复')
super().__init__(s_dim=s_dim,
visual_sources=0,
visual_resolution=0,
a_dim=a_dim,
is_continuous=is_continuous,
**kwargs)
self.n = n
self.i = i
self.ployak = ployak
# self.action_noise = rls.NormalActionNoise(mu=np.zeros(self.a_dim), sigma=1 * np.ones(self.a_dim))
self.action_noise = rls.OrnsteinUhlenbeckActionNoise(
mu=np.zeros(self.a_dim),
sigma=0.2 * np.exp(-self.episode / 10) * np.ones(self.a_dim))
_actor_net = lambda: rls.actor_dpg(self.s_dim, 0, self.a_dim,
hidden_units['actor'])
self.actor_net = _actor_net()
self.actor_target_net = _actor_net()
_q_net = lambda: rls.critic_q_one(
(self.s_dim) * self.n, 0, (self.a_dim) * self.n, hidden_units['q'])
self.critic_net = DoubleQ(_q_net)
self.critic_target_net = DoubleQ(_q_net)
self.update_target_net_weights(
self.actor_target_net.weights + self.critic_target_net.weights,
self.actor_net.weights + self.critic_net.weights)
self.actor_lr, self.critic_lr = map(self.init_lr,
[actor_lr, critic_lr])
self.optimizer_actor, self.optimizer_critic = map(
self.init_optimizer, [self.actor_lr, self.critic_lr])
self.model_recorder(
dict(actor=self.actor_net,
critic_net=self.critic_net,
optimizer_critic=self.optimizer_critic,
optimizer_actor=self.optimizer_actor))
self.recorder.logger.info(self.action_noise)
def _actor_net():
return rls.actor_dpg(self.s_dim, 0, self.a_dim,
hidden_units['actor'])
def _actor_net(): return rls.actor_dpg(self.feat_dim, self.a_dim, hidden_units['actor_continuous']) # self.action_noise = rls.NormalActionNoise(mu=np.zeros(self.a_dim), sigma=1 * np.ones(self.a_dim)) self.action_noise = rls.OrnsteinUhlenbeckActionNoise(mu=np.zeros(self.a_dim), sigma=0.2 * np.ones(self.a_dim))
def _actor_net():
return rls.actor_dpg(self.feat_dim, self.a_dim,
hidden_units['actor_continuous'])
def __init__(
self,
s_dim,
visual_sources,
visual_resolution,
a_dim,
is_continuous,
ployak=0.995,
high_scale=1.0,
reward_scale=1.0,
sample_g_nums=100,
sub_goal_steps=10,
fn_goal_dim=0,
intrinsic_reward_mode='os',
high_batch_size=256,
high_buffer_size=100000,
low_batch_size=8,
low_buffer_size=10000,
high_actor_lr=1.0e-4,
high_critic_lr=1.0e-3,
low_actor_lr=1.0e-4,
low_critic_lr=1.0e-3,
hidden_units={
'high_actor': [64, 64],
'high_critic': [64, 64],
'low_actor': [64, 64],
'low_critic': [64, 64]
},
**kwargs):
assert visual_sources == 0, 'HIRO doesn\'t support visual inputs.'
super().__init__(s_dim=s_dim,
visual_sources=visual_sources,
visual_resolution=visual_resolution,
a_dim=a_dim,
is_continuous=is_continuous,
**kwargs)
self.data_high = ExperienceReplay(high_batch_size, high_buffer_size)
self.data_low = ExperienceReplay(low_batch_size, low_buffer_size)
self.ployak = ployak
self.high_scale = np.array(
high_scale if isinstance(high_scale, list) else [high_scale] *
self.s_dim,
dtype=np.float32)
self.reward_scale = reward_scale
self.fn_goal_dim = fn_goal_dim
self.sample_g_nums = sample_g_nums
self.sub_goal_steps = sub_goal_steps
self.sub_goal_dim = self.s_dim - self.fn_goal_dim
self.high_noise = rls.ClippedNormalActionNoise(
mu=np.zeros(self.sub_goal_dim),
sigma=self.high_scale * np.ones(self.sub_goal_dim),
bound=self.high_scale / 2)
self.low_noise = rls.ClippedNormalActionNoise(mu=np.zeros(self.a_dim),
sigma=1.0 *
np.ones(self.a_dim),
bound=0.5)
_high_actor_net = lambda: rls.actor_dpg(self.s_dim, self.sub_goal_dim,
hidden_units['high_actor'])
if self.is_continuous:
_low_actor_net = lambda: rls.actor_dpg(
self.s_dim + self.sub_goal_dim, self.a_dim, hidden_units[
'low_actor'])
else:
_low_actor_net = lambda: rls.actor_discrete(
self.s_dim + self.sub_goal_dim, self.a_dim, hidden_units[
'low_actor'])
self.gumbel_dist = tfd.Gumbel(0, 1)
self.high_actor = _high_actor_net()
self.high_actor_target = _high_actor_net()
self.low_actor = _low_actor_net()
self.low_actor_target = _low_actor_net()
_high_critic_net = lambda: rls.critic_q_one(
self.s_dim, self.sub_goal_dim, hidden_units['high_critic'])
_low_critic_net = lambda: rls.critic_q_one(
self.s_dim + self.sub_goal_dim, self.a_dim, hidden_units[
'low_critic'])
self.high_critic = DoubleQ(_high_critic_net)
self.high_critic_target = DoubleQ(_high_critic_net)
self.low_critic = DoubleQ(_low_critic_net)
self.low_critic_target = DoubleQ(_low_critic_net)
self.update_target_net_weights(
self.low_actor_target.weights + self.low_critic_target.weights +
self.high_actor_target.weights + self.high_critic_target.weights,
self.low_actor.weights + self.low_critic.weights +
self.high_actor.weights + self.high_critic.weights)
self.low_actor_lr, self.low_critic_lr = map(
self.init_lr, [low_actor_lr, low_critic_lr])
self.high_actor_lr, self.high_critic_lr = map(
self.init_lr, [high_actor_lr, high_critic_lr])
self.low_actor_optimizer, self.low_critic_optimizer = map(
self.init_optimizer, [self.low_actor_lr, self.low_critic_lr])
self.high_actor_optimizer, self.high_critic_optimizer = map(
self.init_optimizer, [self.high_actor_lr, self.high_critic_lr])
self.model_recorder(
dict(high_actor=self.high_actor,
high_critic=self.high_critic,
low_actor=self.low_actor,
low_critic=self.low_critic,
low_actor_optimizer=self.low_actor_optimizer,
low_critic_optimizer=self.low_critic_optimizer,
high_actor_optimizer=self.high_actor_optimizer,
high_critic_optimizer=self.high_critic_optimizer))
self.counts = 0
self._high_s = [[] for _ in range(self.n_agents)]
self._noop_subgoal = np.random.uniform(-self.high_scale,
self.high_scale,
size=(self.n_agents,
self.sub_goal_dim))
self.get_ir = self.generate_ir_func(mode=intrinsic_reward_mode)
def _low_actor_net():
return rls.actor_dpg(self.s_dim + self.sub_goal_dim,
self.a_dim, hidden_units['low_actor'])
def _high_actor_net():
return rls.actor_dpg(self.s_dim, self.sub_goal_dim,
hidden_units['high_actor'])