def init(self, net_dim, state_dim, action_dim):
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.act = Actor(net_dim, state_dim, action_dim).to(self.device)
self.act_target = Actor(net_dim, state_dim, action_dim).to(self.device)
self.cri = Critic(net_dim, state_dim, action_dim).to(self.device)
self.cri_target = Critic(net_dim, state_dim, action_dim).to(self.device)
self.act_optimizer = torch.optim.Adam(self.act.parameters(), lr=self.learning_rate)
self.cri_optimizer = torch.optim.Adam(self.cri.parameters(), lr=self.learning_rate)
def __init__(self, net_dim, state_dim, action_dim, learning_rate=1e-4):
super().__init__()
self.explore_noise = 0.05 # explore noise of action
self.act = Actor(net_dim, state_dim, action_dim).to(self.device)
self.act_target = deepcopy(self.act)
self.cri = Critic(net_dim, 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
}, {
'params': self.cri.parameters(),
'lr': learning_rate
}])
class AgentDDPG(AgentBase):
def __init__(self):
super().__init__()
self.explore_noise = 0.1 # explore noise of action
def init(self, net_dim, state_dim, action_dim):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.act = Actor(net_dim, state_dim, action_dim).to(self.device)
self.act_target = deepcopy(self.act)
self.cri = Critic(net_dim, state_dim, action_dim).to(self.device)
self.cri_target = deepcopy(self.cri)
self.criterion = torch.nn.MSELoss()
self.act_optimizer = torch.optim.Adam(self.act.parameters(),
lr=self.learning_rate)
self.cri_optimizer = torch.optim.Adam(self.cri.parameters(),
lr=self.learning_rate)
def select_action(self, state):
states = torch.as_tensor((state, ),
dtype=torch.float32,
device=self.device).detach_()
action = self.act(states)[0]
action = action + torch.randn_like(action) * self.explore_noise
return action.cpu().numpy()
def update_net(self, buffer, target_step, batch_size, repeat_times):
buffer.update_now_len_before_sample()
obj_critic = obj_actor = None
for _ in range(int(target_step * repeat_times)):
with torch.no_grad():
reward, mask, action, state, next_s = buffer.sample_batch(
batch_size)
next_q = self.cri_target(next_s, self.act_target(next_s))
q_label = reward + mask * next_q
q_value = self.cri(state, action)
obj_critic = self.criterion(q_value, q_label)
self.cri_optimizer.zero_grad()
obj_critic.backward()
self.cri_optimizer.step()
self.soft_update(self.cri_target, self.cri)
q_value_pg = self.act(state) # policy gradient
obj_actor = -self.cri_target(state, q_value_pg).mean()
self.act_optimizer.zero_grad()
obj_actor.backward()
self.act_optimizer.step()
self.soft_update(self.act_target, self.act)
return obj_actor.item(), obj_critic.item()
class AgentDDPG(AgentBase):
def __init__(self, net_dim, state_dim, action_dim, learning_rate=1e-4):
super().__init__()
self.explore_noise = 0.05 # explore noise of action
self.act = Actor(net_dim, state_dim, action_dim).to(self.device)
self.act_target = deepcopy(self.act)
self.cri = Critic(net_dim, 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
}, {
'params': self.cri.parameters(),
'lr': learning_rate
}])
def select_actions(self, states): # states = (state, ...)
states = torch.as_tensor(states,
dtype=torch.float32,
device=self.device)
actions = self.act(states)
actions = (actions +
torch.randn_like(actions) * self.explore_noise).clamp(
-1, 1)
return actions.detach().cpu().numpy()
def update_net(self, buffer, max_step, batch_size, repeat_times):
buffer.update__now_len__before_sample()
obj_critic = obj_actor = None # just for print return
for _ in range(int(max_step * repeat_times)):
with torch.no_grad():
reward, mask, action, state, next_s = buffer.random_sample(
batch_size)
next_q = self.cri_target(next_s, self.act_target(next_s))
q_label = reward + mask * next_q
q_value = self.cri(state, action)
obj_critic = self.criterion(q_value, q_label)
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()
soft_target_update(self.cri_target, self.cri)
soft_target_update(self.act_target, self.act)
return obj_actor.item(), obj_critic.item()
