class Agent(object):
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim
self.__device = device
self.__gamma = gamma
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
self.__policy = DQN(action_dim, device).to(device)
self.__target = DQN(action_dim, device).to(device)
if restore is None:
self.__policy.apply(DQN.init_weights)
else:
self.__policy.load_state_dict(torch.load(restore))
self.__target.load_state_dict(self.__policy.state_dict())
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
def run(self,
state: TensorStack4,
training: bool = False,
preaction: int = 0) -> int:
"""run suggests an action for the given state."""
if training:
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps:
with torch.no_grad():
return self.__policy(state).max(1).indices.item()
if self.__r.random() > 0.15:
return preaction
return self.__r.randint(0, self.__action_dim - 1)
def learn(self, memory: ReplayMemory, batch_size: int) -> float:
"""learn trains the value network via TD-learning."""
state_batch, action_batch, reward_batch, next_batch, done_batch = \
memory.sample(batch_size)
values = self.__policy(state_batch.float()).gather(1, action_batch)
values_next = self.__target(next_batch.float()).max(1).values.detach()
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch
loss = F.smooth_l1_loss(values, expected)
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters():
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object): #就是那个反弹平板模型
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
restore: Optional[str] = None, #restore默认设为None
) -> None:
self.__action_dim = action_dim #设置模型初始参数
self.__device = device
self.__gamma = gamma
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
self.__policy = DQN(action_dim, device).to(device) #可以设置运行在cpu还是gpu上
self.__target = DQN(action_dim, device).to(device)
if restore is None:
self.__policy.apply(DQN.init_weights) #采用dqn初始权重
else:
self.__policy.load_state_dict(
torch.load(restore)) #否则更新为restore中的权重
self.__target.load_state_dict(self.__policy.state_dict())
self.__optimizer = optim.Adam( #优化器
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
# 采用epsilon-greedy方法选择action
def run(self, state: TensorStack4, training: bool = False) -> int: #运行
"""run suggests an action for the given state."""
if training:
#利用eps_start/eps_final这些动态衰减epsilon
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay #这是啥意思？？貌似是为了调整学习率
self.__eps = max(self.__eps, self.__eps_final)
#有一定的概率选择使函数值最大的action，在当前网络policy_network中得到
if self.__r.random() > self.__eps:
with torch.no_grad():
return self.__policy(state).max(1).indices.item()
#否则随机选择action
return self.__r.randint(0, self.__action_dim - 1)
#从memory中提取state action reward next来训练网络
def learn(self, memory: ReplayMemory, batch_size: int) -> float:
"""learn trains the value network via TD-learning."""
# 从replay buffer当中采样！！用于更新policy网络
state_batch, action_batch, reward_batch, next_batch, done_batch = \
memory.sample(batch_size) #将所有变量转为张量。
#使用行为网络计算值函数 Q_j
values = self.__policy(state_batch.float()).gather(
1, action_batch) # Q_j对应有state_batch和action_batch
#使用目标网络计算 Q_{j+1}并计算 expected = r_{j+1} + max(a') Q_{j+1}
#其中(1-done_batch)用于判断是否terminal，如果是就退化到expected = r_{j+1}
#这里相当于q-learning中的更新公式的一部分。在target网络中计算Q值。
values_next = self.__target(next_batch.float()).max(1).values.detach()
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch
#根据目标函数 （Q_j - expected)^2来梯度下降
loss = F.smooth_l1_loss(values, expected)
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters():
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
#同步target网络和policy网络，即目标和行为网络
def sync(self) -> None: #
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
#保存policy network
def save(self, path: str) -> None: #保存结果
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object):
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim
self.__device = device
self.__gamma = gamma
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
self.__policy = DQN(action_dim, device).to(device)
self.__target = DQN(action_dim, device).to(device)
if restore is None:
self.__policy.apply(DQN.init_weights)
else:
self.__policy.load_state_dict(torch.load(restore))
self.__target.load_state_dict(self.__policy.state_dict())
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
def run(self, state: TensorStack4, training: bool = False) -> int:
"""run suggests an action for the given state."""
if training:
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps:
with torch.no_grad():
action = self.__policy(state).max(1).indices.item()
else:
action = self.__r.randint(0, self.__action_dim - 1)
value_this = self.__policy(state)[0][action]
return action, value_this #行动和Q值
#return self.__r.randint(0, self.__action_dim - 1)
def get_target_value(self, state):
value_next = self.__target(state).max(1).indices.item()
return value_next
def learn(self, memory, batch_size: int, c) -> float: #####
"""learn trains the value network via TD-learning."""
if (c == 0):
state_batch, action_batch, reward_batch, next_batch, done_batch = memory.sample(
batch_size)
else:
state_batch, action_batch, reward_batch, next_batch, done_batch ,idx_batch= \
memory.sample(batch_size) ####
values = self.__policy(state_batch.float()).gather(
1, action_batch) #每一列按action_batch取元素
values_next = self.__target(
next_batch.float()).max(1).values.detach() #最大的作为下一个的value
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch
#Loss=values-expected
if (c == 0): #####
loss = F.smooth_l1_loss(values, expected)
else:
loss_batch = F.smooth_l1_loss(values, expected,
reduction='none') #TD error
loss = torch.mean(loss_batch, dim=0)
#loss.requires_grad = True
memory.update(loss_batch.detach(), idx_batch)
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters(): #把参数加紧到[-1,1],原地修改
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object): # 智能体的配置
def __init__(
self,
action_dim: int, # 3
device: TorchDevice, # cuda
gamma: float, # 0.99
seed: int,
eps_start: float, # 1
eps_final: float, # 0.1
eps_decay: float, # 10000
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim # 3
self.__device = device
self.__gamma = gamma # 0.99
self.__eps_start = eps_start # 1
self.__eps_final = eps_final # 0.1
self.__eps_decay = eps_decay # 1e6
self.__eps = eps_start # 1
self.__r = random.Random()
self.__r.seed(seed)
self.__policy = DQN(action_dim, device).to(device) # 策略DQN
self.__target = DQN(action_dim,
device).to(device) # 目标DQN，减少目标计算与当前值的相关性
if restore is None: self.__policy.apply(DQN.init_weights)
else: self.__policy.load_state_dict(torch.load(restore))
self.__target.load_state_dict(
self.__policy.state_dict()) # 将策略网络中的权重同步到目标网络
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
) # 优化器
self.__target.eval() # 验证模式
def run(self, state: TensorStack4, training: bool = False) -> int:
"""run suggests an action for the given state."""
if training: # 修改eps，逐渐降低
self.__eps -= (self.__eps_start -
self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps: # 1-eps概率取最大值
with torch.no_grad():
return self.__policy(state).max(1).indices.item()
return self.__r.randint(0, self.__action_dim - 1) # eps概率随机
def learn(self, memory: ReplayMemory, batch_size: int) -> float: # 训练
"""learn trains the value network via TD-learning."""
state_batch, action_batch, reward_batch, next_batch, done_batch = memory.sample(
batch_size) # 在经验池中选取一组transition样本集（minibatch）
values = self.__policy(state_batch.float()).gather(
1, action_batch) # DQN输出y
values_next = self.__target(
next_batch.float()).max(1).values.detach() # max_a(Q(S',a))
expected = (self.__gamma * values_next.unsqueeze(1)) * (
1. - done_batch) + reward_batch # Q-Learning计算Q(S,A)
loss = F.smooth_l1_loss(values, expected) # 计算误差
# 更新网络参数三部曲
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters():
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item() # 返回误差
def sync(self) -> None: # 将策略网络中的权重同步到目标网络
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None: # 保存模型
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object): # 代理
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
isdueling: bool = False,
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim # 动作维度
self.__device = device # 设备
self.__gamma = gamma # 衰减因子
self.__eps_start = eps_start # eps-greedy参数的初始值
self.__eps_final = eps_final # eps-greedy参数的最终值
self.__eps_decay = eps_decay # eps-greedy参数的衰减率
self.__eps = eps_start
self.__r = random.Random() #随机浮点数
self.__r.seed(seed) # 随机数种子
###修改项
if isdueling: # 使用DuelingDQN网络
self.__policy = DuelingDQN(action_dim, device).to(device) # 值函数网络
self.__target = DuelingDQN(action_dim, device).to(device) # target网络
else:
self.__policy = DQN(action_dim, device).to(device) # 值函数网络
self.__target = DQN(action_dim, device).to(device) # target网络
if restore is None:
if isdueling:
self.__policy.apply(DuelingDQN.init_weights) # 初始化权重
else:
self.__policy.apply(DQN.init_weights) # 初始化权重
###修改项
else:
self.__policy.load_state_dict(torch.load(restore)) # 将restore的数据加载到网络中
self.__target.load_state_dict(self.__policy.state_dict()) # 将policy参数赋给target，此时两个网络相同
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
def run(self, state: TensorStack4, training: bool = False) -> int: #e-greedy选择行动
"""run suggests an action for the given state."""
if training: # 线性衰减eps
self.__eps -= (self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps: # 产生随机数>eps，选择使Q最大的动作
with torch.no_grad():
action = self.__policy(state).max(1).indices.item()
else:
action = self.__r.randint(0, self.__action_dim - 1)
value_this = self.__policy(state)[0][action]
return action, value_this # 行动和Q值
def get_target_value(self, state):
value_next = self.__target(state).max(1).indices.item()
return value_next
def learn(self, memory: ReplayMemory, batch_size: int, choice: int) -> float:
"""learn trains the value network via TD-learning."""
##修改项
if (choice == 0): # 普通memory
state_batch, action_batch, reward_batch, next_batch, done_batch = memory.sample(batch_size)
else: # PERmemory
state_batch, action_batch, reward_batch, next_batch, done_batch, idx_batch = \
memory.sample(batch_size) ####
values = self.__policy(state_batch.float()).gather(1, action_batch) # 每一列按action_batch取元素
values_next = self.__target(next_batch.float()).max(1).values.detach() # 最大的作为下一个的value
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch
# Loss=values-expected
if (choice == 0): #####
loss = F.smooth_l1_loss(values, expected)
else: # PERmemory
loss_batch = F.smooth_l1_loss(values, expected, reduction='none') # TD error
loss = torch.mean(loss_batch, dim=0)
# loss.requires_grad = True
memory.update(loss_batch.detach(), idx_batch)
##修改项
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters(): # 把参数加紧到[-1,1],原地修改
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object):
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim
self.__device = device
self.__gamma = gamma
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
# 将所有最开始读取数据时的tensor变量copy一份到device所指定的GPU上去，之后的运算都在GPU上进行
self.__policy = DQN(action_dim, device).to(device) # policy network
self.__target = DQN(action_dim, device).to(device) # target network
if restore is None:
self.__policy.apply(DQN.init_weights) # policy自定义参数初始化方式
else:
self.__policy.load_state_dict(
torch.load(restore)) # policy加载之前学习到的参数
self.__target.load_state_dict(
self.__policy.state_dict()) # target拷贝policy的参数
self.__optimizer = optim.Adam( # 优化器采用Adam
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval() # 将模型转变为evaluation（测试）模式，这样就可以排除BN和Dropout对测试的干扰
# epsilon-greedy
def run(self, state: TensorStack4, training: bool = False) -> int:
"""run suggests an action for the given state."""
if training:
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps:
with torch.no_grad():
return self.__policy(state).max(1).indices.item()
return self.__r.randint(0, self.__action_dim - 1)
def learn(self, memory: Experience, step: int) -> float:
"""learn trains the value network via TD-learning."""
state_batch, action_batch, reward_batch, next_batch, done_batch, w, rank_e_id = \
memory.sample(step) # 随机取样 state是5帧的前4帧 next是5帧的后4帧
values = self.__policy(state_batch.float()).gather(1, action_batch)
# values_next = self.__target(next_batch.float()).max(1).values.detach() # 这里还是nature dqn 没有用ddqn 虽都是双网络
values_next = self.__target(next_batch.float()).gather(
1,
self.__policy(next_batch.float()).max(1).indices.unsqueeze(
1)).detach() # 改成ddqn
reward_batch[action_batch == 0] += 0.1 # stable reward
expected = (self.__gamma * values_next) * \
(1. - done_batch) + reward_batch # 如果done则是r（考虑t时刻done，没有t+1时刻），否则是r + gamma * max Q
td_error = (expected - values).detach()
memory.update_priority(rank_e_id, td_error.cpu().numpy())
values = values.mul(w)
expected = expected.mul(w)
loss = F.smooth_l1_loss(values, expected) # smooth l1损失
self.__optimizer.zero_grad() # 将模型的参数梯度初始化为0
loss.backward() # 计算梯度，存到__policy.parameters.grad()中
for param in self.__policy.parameters():
param.grad.data.clamp_(-1, 1) # 固定所有梯度为[-1, 1]
self.__optimizer.step() # 做一步最优化
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object):
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
restore: Optional[str] = None,
) -> None:
self.__action_dim = action_dim #动作可选维数
self.__device = device #设备
self.__gamma = gamma #γ值
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
self.__policy = DQN(action_dim, device).to(device) # policy网络
self.__target = DQN(action_dim, device).to(device) # target网络
if restore is None:
self.__policy.apply(DQN.init_weights)
else:
self.__policy.load_state_dict(torch.load(restore))
self.__target.load_state_dict(self.__policy.state_dict())
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
def run(self,
state: TensorStack4,
training: bool = False) -> int: # 返回action
"""run suggests an action for the given state."""
if training:
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps:
with torch.no_grad():
return self.__policy(state).max(1).indices.item()
return self.__r.randint(0, self.__action_dim - 1)
def learn(self, memory: ReplayMemory, batch_size: int) -> float: # 返回vlue
"""learn trains the value network via TD-learning."""
state_batch, action_batch, reward_batch, next_batch, done_batch = \
memory.sample(batch_size)#随机选取一个样本
# SGD优化的基本要求之一是训练数据是独立且均匀分布的
# 当Agent与环境交互时，经验元组的序列可以高度相关,所以要打乱采样
#将样本送入学习
values = self.__policy(state_batch.float()).gather(
1, action_batch) #Q表：value=Q(s,a)
##########dueling DQN修改思路:拆分Q(s,a)=V(s)+A(s,a)，其中V(s)为状态s本身的价值，A(s,a)为动作a的价值##########
##########V(s)是一个标量，A(s,a)是一个向量。在相加时V(s)会自动复制到与A(s,a)维度一致##########
values_next = self.__target(
next_batch.float()).max(1).values.detach() #Q'表，但是具体参数不清楚
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch #当完成了（done=1），y_j=r_j;否则y_j=r_j+q()见论文算法
loss = F.smooth_l1_loss(values, expected) # 损失函数
self.__optimizer.zero_grad()
loss.backward()
for param in self.__policy.parameters():
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
class Agent(object):
def __init__(
self,
action_dim: int,
device: TorchDevice,
gamma: float,
seed: int,
eps_start: float,
eps_final: float,
eps_decay: float,
dueling: bool,
restore: Optional[str] = None,
stable_arg = 0.1,
) -> None:
self.__action_dim = action_dim
self.__device = device
self.__gamma = gamma
self.__eps_start = eps_start
self.__eps_final = eps_final
self.__eps_decay = eps_decay
self.__eps = eps_start
self.__r = random.Random()
self.__r.seed(seed)
self.__stable_arg = stable_arg
if dueling:
self.__policy = DuelingDQN(action_dim, device).to(device)
self.__target = DuelingDQN(action_dim, device).to(device)
else:
self.__policy = DQN(action_dim, device).to(device)
self.__target = DQN(action_dim, device).to(device)
if restore is None:
self.__policy.apply(DQN.init_weights)
else:
#if dueling:
#self.__policy.Convs_load(restore)
#else:
self.__policy.load_state_dict(torch.load(restore))
self.__target.load_state_dict(self.__policy.state_dict())
self.__optimizer = optim.Adam(
self.__policy.parameters(),
lr=0.0000625,
eps=1.5e-4,
)
self.__target.eval()
def run(self, state: TensorStack4, training: bool = False) -> int: #epsilon greedy policy
"""run suggests an action for the given state."""
if training:
self.__eps -= \
(self.__eps_start - self.__eps_final) / self.__eps_decay
self.__eps = max(self.__eps, self.__eps_final)
if self.__r.random() > self.__eps:
with torch.no_grad():
action = self.__policy(state).max(1).indices.item()
else:
action = self.__r.randint(0, self.__action_dim - 1)
value_this = self.__policy(state)[0][action]
return action, value_this
def get_target_value(self, state):
value_next = self.__target(state).max(1).indices.item()
return value_next
def learn(self, memory: ReplayMemory, batch_size: int) -> float:
"""learn trains the value network via TD-learning.应该是Q network """
state_batch, action_batch, reward_batch, \
next_batch, done_batch, idx_batch = memory.sample(batch_size)
values = self.__policy(state_batch.float()).gather(1, action_batch)
values_next = self.__target(next_batch.float()).max(1).values.detach()
expected = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done_batch) + reward_batch #TD target
loss_batch = F.smooth_l1_loss(values, expected, reduce=False) #TD error
loss = torch.mean(loss_batch, dim=0)
loss.requires_grad = True
memory.update(loss_batch.detach(), idx_batch)
self.__optimizer.zero_grad()
loss.backward() #backward
for param in self.__policy.parameters():
if param.grad is not None:#grad clamp to (-1,1)
param.grad.data.clamp_(-1, 1)
self.__optimizer.step() #update
return loss.item()
def stable_learn(self, folded_state, action, reward, done):
'''learn stable and q_value'''
state = folded_state[:4]
state_next = folded_state[1:]
value_next = agent.get_target_value(state_next)
value_now = self.__policy(state.float()).gather(1, action)
td_target = (self.__gamma * values_next.unsqueeze(1)) * \
(1. - done) + reward
td_error = F.smooth_l1_loss(value_now, td_target)
stable_loss = torch.zeros(1).float()
for i in [1, 2]:
for j in [i-1, i+1]:
stable_loss += 1*(action[j]*action[i]==2)
stable_loss -= 1*(action[1]*action[2]==2)
stable_loss.requires_grad = True
loss = td_error + self.stable_arg*stable_loss
self.__optimizer.zero_grad()
loss.backward() #backward
for param in self.__policy.parameters():
if param.grad is not None:#grad clamp to (-1,1)
param.grad.data.clamp_(-1, 1)
self.__optimizer.step()
return loss.item()
def sync(self) -> None:
"""sync synchronizes the weights from the policy network to the target
network."""
self.__target.load_state_dict(self.__policy.state_dict())
def save(self, path: str) -> None:
"""save saves the state dict of the policy network."""
torch.save(self.__policy.state_dict(), path)
def load(self, path: str) -> None:
"""save saves the state dict of the policy network."""
self.__policy.load_state_dict(path)
self.__target.load_state_dict(path)