class AgentTD3(AgentDDPG):
def __init__(self, net_dim, state_dim, action_dim, learning_rate=1e-4):
super().__init__(net_dim, state_dim, action_dim, learning_rate)
self.explore_noise = 0.1 # standard deviation of explore noise
self.policy_noise = 0.2 # standard deviation of policy noise
self.update_freq = 2 # delay update frequency, for soft target update
self.cri = CriticTwin(net_dim, state_dim, action_dim).to(self.device)
self.cri_target = deepcopy(self.cri)
self.optimizer = torch.optim.Adam([{
'params': self.act.parameters(),
'lr': learning_rate
}, {
'params': self.cri.parameters(),
'lr': learning_rate
}])
def update_policy(self, buffer, max_step, batch_size, repeat_times):
buffer.update__now_len__before_sample()
obj_critic = obj_actor = None
for i in range(int(max_step * repeat_times)):
with torch.no_grad():
reward, mask, action, state, next_s = buffer.random_sample(
batch_size)
next_a = self.act_target.get_action(
next_s, self.policy_noise) # policy noise
next_q = torch.min(*self.cri_target.get__q1_q2(
next_s, next_a)) # twin critics
q_label = reward + mask * next_q
q1, q2 = self.cri.get__q1_q2(state, action)
obj_critic = self.criterion(q1, q_label) + self.criterion(
q2, q_label) # twin critics
q_value_pg = self.act(state) # policy gradient
obj_actor = -self.cri_target(state, q_value_pg).mean()
obj_united = obj_actor + obj_critic # objective
self.optimizer.zero_grad()
obj_united.backward()
self.optimizer.step()
if i % self.update_freq == 0: # delay update
soft_target_update(self.cri_target, self.cri)
soft_target_update(self.act_target, self.act)
return obj_actor.item(), obj_critic.item() / 2
class AgentSAC(AgentBase):
def __init__(self, net_dim, state_dim, action_dim, learning_rate=1e-4):
super().__init__()
self.target_entropy = np.log(action_dim)
self.alpha_log = torch.tensor(
(-np.log(action_dim) * np.e, ),
dtype=torch.float32,
requires_grad=True,
device=self.device) # trainable parameter
self.act = ActorSAC(net_dim, state_dim, action_dim).to(self.device)
self.act_target = deepcopy(self.act)
self.cri = CriticTwin(
int(net_dim * 1.25),
state_dim,
action_dim,
).to(self.device)
self.cri_target = deepcopy(self.cri)
self.criterion = torch.nn.SmoothL1Loss()
self.optimizer = torch.optim.Adam([{
'params': self.act.parameters(),
'lr': learning_rate * 0.75
}, {
'params': self.cri.parameters(),
'lr': learning_rate * 1.25
}, {
'params': (self.alpha_log, ),
'lr': learning_rate
}])
def select_actions(self, states): # states = (state, ...)
states = torch.as_tensor(states,
dtype=torch.float32,
device=self.device)
actions = self.act.get_action(states)
return actions.detach().cpu().numpy()
def update_policy(self, buffer, max_step, batch_size, repeat_times):
buffer.update__now_len__before_sample()
alpha = self.alpha_log.exp().detach()
obj_actor = obj_critic = None
for _ in range(int(max_step * repeat_times)):
with torch.no_grad():
reward, mask, action, state, next_s = buffer.random_sample(
batch_size)
next_a, next_log_prob = self.act_target.get__action__log_prob(
next_s)
next_q = torch.min(*self.cri_target.get__q1_q2(next_s, next_a))
q_label = reward + mask * (next_q + next_log_prob * alpha)
q1, q2 = self.cri.get__q1_q2(state, action)
obj_critic = self.criterion(q1, q_label) + self.criterion(
q2, q_label)
action_pg, log_prob = self.act.get__action__log_prob(
state) # policy gradient
obj_alpha = (self.alpha_log *
(log_prob - self.target_entropy).detach()).mean()
alpha = self.alpha_log.exp().detach()
with torch.no_grad():
self.alpha_log[:] = self.alpha_log.clamp(-16, 2)
obj_actor = -(
torch.min(*self.cri_target.get__q1_q2(state, action_pg)) +
log_prob * alpha).mean()
obj_united = obj_critic + obj_alpha + obj_actor
self.optimizer.zero_grad()
obj_united.backward()
self.optimizer.step()
soft_target_update(self.cri_target, self.cri)
soft_target_update(self.act_target, self.act)
# return obj_actor.item(), obj_critic.item()
return alpha.item(), obj_critic.item()