一、DDPG简介
PPO这类算法都是在线策略算法,样本效率(sample efficiency)较低。像DQN算法,是直接估计最优价值函数,可以做离线策略学习,但是它只能处理动作空间有限的环境。
吸收DQN的优点,同时弥补PPO这类算法的缺陷,DDPG(deep deterministic policy gradient)就顺应而生。它构造一个确定性策略,用梯度上升的方法来最大化值。
DDPG 也属于一种 Actor-Critic 算法。 REINFORCE、TRPO 和 PPO 学习随机性策略(Actor
输出action的分布进行采样),而 DDPG 则学习一个确定性策略(Actor
直接输出action)。
二、 Pytorch实践
策略网络(Actor
)直接输出确定性action
class policyNet(nn.Module):
"""
return continuity action
"""
def __init__(self, state_dim: int, hidden_layers_dim: typ.List, action_dim: int, action_bound: float=1.0):
super(policyNet, self).__init__()
self.action_bound = action_bound
self.features = nn.ModuleList()
for idx, h in enumerate(hidden_layers_dim):
self.features.append(nn.ModuleDict({
'linear': nn.Linear(hidden_layers_dim[idx-1] if idx else state_dim, h),
'linear_action': nn.ReLU(inplace=True)
}))
self.fc_out = nn.Linear(hidden_layers_dim[-1], action_dim)
def forward(self, x):
for layer in self.features:
x = layer['linear_action'](layer['linear'](x))
return torch.tanh(self.fc_out(x)) * self.action_bound
DQN输入state返回所有action的价值,再用max选取action
DDPG的ValueNet直接以state和action为入参,输出价值
class valueNet(nn.Module):
def __init__(self, state_action_dim: int, hidden_layers_dim: typ.List):
super(valueNet, self).__init__()
self.features = nn.ModuleList()
for idx, h in enumerate(hidden_layers_dim):
self.features.append(nn.ModuleDict({
'linear': nn.Linear(hidden_layers_dim[idx-1] if idx else state_action_dim, h),
'linear_activation': nn.ReLU(inplace=True)
}))
self.head = nn.Linear(hidden_layers_dim[-1] , 1)
def forward(self, state, action):
x = torch.cat([state, action], dim=1).float() # 拼接状态和动作
for layer in self.features:
x = layer['linear_activation'](layer['linear'](x))
return self.head(x)
2.1 构建智能体
同时在action输出的时候增加一些噪声act.detach().numpy()[0] + self.sigma * np.random.rand(self.action_dim)
class DDPG:
def __init__(self,
state_dim: int,
hidden_layers_dim: typ.List,
action_dim: int,
actor_lr: float,
critic_lr: float,
gamma: float,
DDPG_kwargs: typ.Dict,
device: torch.device
):
self.actor = policyNet(state_dim, hidden_layers_dim, action_dim, action_bound = DDPG_kwargs['action_bound'])
self.critic = valueNet(state_dim + action_dim, hidden_layers_dim)
self.target_actor = copy.deepcopy(self.actor)
self.target_critic = copy.deepcopy(self.critic)
self.actor.to(device)
self.critic.to(device)
self.target_actor.to(device)
self.target_critic.to(device)
self.actor_opt = torch.optim.Adam(self.actor.parameters(), lr=actor_lr)
self.critic_opt = torch.optim.Adam(self.critic.parameters(), lr=critic_lr)
self.gamma = gamma
self.device = device
self.count = 0
# soft update parameters
self.tau = DDPG_kwargs.get('tau', 0.8)
self.action_dim = action_dim
# Normal sigma
self.sigma = DDPG_kwargs.get('sigma', 0.1)
def policy(self, state):
state = torch.FloatTensor([state]).to(self.device)
act = self.actor(state)
return act.detach().numpy()[0] + self.sigma * np.random.rand(self.action_dim)
def soft_update(self, net, target_net):
for param_target, param in zip(target_net.parameters(), net.parameters()):
param_target.data.copy_(
param_target.data * (1 - self.tau) + param.data * self.tau
)
def update(self, samples):
self.count += 1
state, action, reward, next_state, done = zip(*samples)
state = torch.FloatTensor(state).to(self.device)
action = torch.tensor(action).to(self.device)
reward = torch.tensor(reward).view(-1, 1).to(self.device)
reward = (reward + 10.0) / 10.0 # 对奖励进行修改,方便训练
next_state = torch.FloatTensor(next_state).to(self.device)
done = torch.FloatTensor(done).view(-1, 1).to(self.device)
target_action = self.target_actor(state)
next_q_value = self.target_critic(next_state, target_action)
q_targets = reward + self.gamma * next_q_value * ( 1.0 - done )
critic_loss = torch.mean(F.mse_loss(self.critic(state, action).float(), q_targets.float()))
self.critic_opt.zero_grad()
critic_loss.backward()
self.critic_opt.step()
# 计算采样的策略梯度,以此更新当前 Actor 网络
ac_action = self.actor(state)
actor_loss = -torch.mean(self.critic(state, ac_action))
self.actor_opt.zero_grad()
actor_loss.backward()
self.actor_opt.step()
self.soft_update(self.actor, self.target_actor)
self.soft_update(self.critic, self.target_critic)
2.2 智能体训练
注意设置tau
要设置小一些
def train_agent(env, cfg):
ac_agent = DDPG(
state_dim=cfg.state_dim,
hidden_layers_dim=cfg.hidden_layers_dim,
action_dim=cfg.action_dim,
actor_lr=cfg.actor_lr,
critic_lr=cfg.critic_lr,
gamma=cfg.gamma,
DDPG_kwargs=cfg.DDPG_kwargs,
device=cfg.device
)
tq_bar = tqdm(range(cfg.num_episode))
rewards_list = []
now_reward = 0
bf_reward = -np.inf
buffer_ = replayBuffer(cfg.buffer_size)
for i in tq_bar:
tq_bar.set_description(f'Episode [ {i+1} / {cfg.num_episode} ]')
s, _ = env.reset()
done = False
episode_rewards = 0
steps = 0
while not done:
a = ac_agent.policy(s)
n_s, r, done, _, _ = env.step(a)
buffer_.add(s, a, r, n_s, done)
s = n_s
episode_rewards += r
steps += 1
# buffer update
if len(buffer_) > cfg.minimal_size:
samples = buffer_.sample(cfg.batch_size)
ac_agent.update(samples)
if (episode_rewards >= cfg.max_episode_rewards) or (steps >= cfg.max_episode_steps):
break
rewards_list.append(episode_rewards)
now_reward = np.mean(rewards_list[-10:])
if (bf_reward < now_reward) and (i >= 10):
torch.save(ac_agent.actor.state_dict(), cfg.save_path)
bf_reward = now_reward
tq_bar.set_postfix({'lastMeanRewards': f'{now_reward:.2f}', 'BEST': f'{bf_reward:.2f}'})
env.close()
return ac_agent
class Config:
num_episode = 200
state_dim = None
hidden_layers_dim = [ 128, 64 ]
action_dim = 20
actor_lr = 3e-5
critic_lr = 5e-4
DDPG_kwargs = {
'tau': 0.05, # soft update parameters
'sigma': 0.005, # noise
'action_bound': 2.0
}
gamma = 0.9
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
buffer_size = 10240
minimal_size = 1024
batch_size = 256
save_path = r'D:\tmp\DDPG_ac_model.ckpt'
# 回合停止控制
max_episode_rewards = 204800
max_episode_steps = 240
def __init__(self, env):
self.state_dim = env.observation_space.shape[0]
try:
self.action_dim = env.action_space.n
except Exception as e:
self.action_dim = env.action_space.shape[0]
print(f'device={self.device} | env={str(env)}')
if __name__ == '__main__':
print('=='*35)
print('Training Pendulum-v1')
env = gym.make('Pendulum-v1')
cfg = Config(env)
ac_agent = train_agent(env, cfg)
2.3 训练出的智能体观测
最后将训练的最好的网络拿出来进行观察
ac_agent.actor.load_state_dict(torch.load(cfg.save_path))
play(gym.make('Pendulum-v1', render_mode="human"), ac_agent, cfg)