策略网络为批处理状态和单个状态返回不同的输出

Question

我正在实施应用于 CartPole-V0 openAI 健身房环境的 REINFORCE。我正在尝试相同的 2 种不同实现，我无法解决的问题如下：

将单个状态传递到策略网络后，我得到一个大小为 2 的输出张量，其中包含 2 个动作的动作概率。但是，当我将“一批状态”传递给策略网络以计算所有状态的输出动作概率时，我获得的值与每个状态单独传递给网络时的值非常不同。

谁能帮我理解这个问题？

我的代码如下：（注意：这不是完整的强化算法——我知道我需要根据概率计算损失。但我试图理解计算中的差异在继续之前，我认为这两个概率应该是相同的。）

# architecture of the Policy Network
class PolicyNetwork(nn.Module):
    def __init__(self, state_dim, n_actions):
        super().__init__()
        self.n_actions = n_actions
        self.model = nn.Sequential(
            nn.Linear(state_dim, 64),
            nn.ReLU(),
            nn.Linear(64, n_actions),
            nn.Softmax(dim=0)
        ).float()

    def forward(self, X):
        return self.model(X)


def train_reinforce_agent(env, episode_length, max_episodes, gamma, visualize_step, learning_rate=0.003):

    # define the parametric model for the Policy: this is an instantiation of the PolicyNetwork class
    model = PolicyNetwork(env.observation_space.shape[0], env.action_space.n)
    # define the optimizer for updating the weights of the Policy Network
    optimizer = optim.Adam(model.parameters(), lr=learning_rate)


    # hyperparameters of the reinforce agent
    EPISODE_LENGTH = episode_length
    MAX_EPISODES = max_episodes
    GAMMA = gamma
    VISUALIZE_STEP = max(1, visualize_step)
    score = []


    for episode in range(MAX_EPISODES):
        # reset the environment
        curr_state = env.reset()
        done = False
        episode_t = []


        # rollout an entire episode from the Policy Network
        pred_vals = []
        for t in range(EPISODE_LENGTH):
            act_prob = model(torch.from_numpy(curr_state).float())
            pred_vals.append(act_prob)
            action = np.random.choice(np.array(list(range(env.action_space.n))), p=act_prob.data.numpy())
            prev_state = curr_state
            curr_state, _, done, info = env.step(action)
            episode_t.append((prev_state, action, t+1))
            if done:
                break
        score.append(len(episode_t))
        # reward_batch = torch.Tensor([r for (s,a,r) in episode_t]).flip(dims=(0,))
        reward_batch = torch.Tensor([r for (s, a, r) in episode_t])


        # compute the return for every state-action pair from the rewards at every time-step
        batch_Gvals = []
        for i in range(len(episode_t)):
            new_Gval = 0
            power = 0
            for j in range(i, len(episode_t)):
                new_Gval = new_Gval + ((GAMMA ** power) * reward_batch[j]).numpy()
            power += 1
            batch_Gvals.append(new_Gval)



        # normalize the returns for the batch
        expected_returns_batch = torch.FloatTensor(batch_Gvals)
        if torch.is_nonzero(expected_returns_batch.max()):
            expected_returns_batch /= expected_returns_batch.max()



        # batch the states, actions, prob after the episode
        state_batch = torch.Tensor([s for (s,a,r) in episode_t])
        print("State batch:", state_batch)
        all_states = [s for (s,a,r) in episode_t]
        print("All states:", all_states)
        action_batch = torch.Tensor([a for (s,a,r) in episode_t])
        pred_batch_v1 = model(state_batch)
        pred_batch_v2 = torch.stack(pred_vals)
        print("Batched state pred_vals:", pred_batch_v1)
        print("Individual state pred_vals:", pred_batch_v2) ### Why is this different from the above predicted values??

我传递环境的主要功能是：

def main():
    env = gym.make('CartPole-v0')
    # train a REINFORCE-agent to learn the optimal policy
    episode_length = 500
    n_episodes = 500
    gamma = 0.99
    vis_steps = 50
    train_reinforce_agent(env, episode_length, n_episodes, gamma, vis_steps)

Answer 1

在你的策略中，你的 Softmax 超过暗淡的 0。这规范化了你的批次中每个动作的概率。您想通过 dim=1.

跨操作执行此操作

策略网络为批处理状态和单个状态返回不同的输出

Policy Network returning different outputs for batched states and individual states

reinforcement-learning

neural-network

deep-learning

openai-gym

pytorch