理解梯度策略推导

Question

我正在尝试从其原始资源Andrej Karpathy blog 重新创建非常简单的策略梯度示例。在那篇文章中，您将找到带有权重列表和 Softmax 激活列表的 CartPole 和 Policy Gradient 示例。这是我重新创建的非常简单的 CartPole 策略梯度示例，效果很好。

import gym
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import PolynomialFeatures
import copy

NUM_EPISODES = 4000
LEARNING_RATE = 0.000025
GAMMA = 0.99


# noinspection PyMethodMayBeStatic
class Agent:
    def __init__(self):
        self.poly = PolynomialFeatures(1)
        self.w = np.random.rand(5, 2)

    def policy(self, state):
        z = state.dot(self.w)
        exp = np.exp(z)
        return exp/np.sum(exp)

    def __softmax_grad(self, softmax):
        s = softmax.reshape(-1,1)
        return np.diagflat(s) - np.dot(s, s.T)

    def grad(self, probs, action, state):
        dsoftmax = self.__softmax_grad(probs)[action,:]
        dlog = dsoftmax / probs[0,action]
        grad = state.T.dot(dlog[None,:])
        return grad

    def update_with(self, grads, rewards):

        for i in range(len(grads)):
            # Loop through everything that happend in the episode
            # and update towards the log policy gradient times **FUTURE** reward

            total_grad_effect = 0
            for t, r in enumerate(rewards[i:]):
                total_grad_effect += r * (GAMMA ** r)
            self.w += LEARNING_RATE * grads[i] * total_grad_effect
            print("Grads update: " + str(np.sum(grads[i])))



def main(argv):
    env = gym.make('CartPole-v0')
    np.random.seed(1)

    agent = Agent()
    complete_scores = []

    for e in range(NUM_EPISODES):
        state = env.reset()[None, :]
        state = agent.poly.fit_transform(state)

        rewards = []
        grads = []
        score = 0

        while True:

            probs = agent.policy(state)
            action_space = env.action_space.n
            action = np.random.choice(action_space, p=probs[0])

            next_state, reward, done,_ = env.step(action)
            next_state = next_state[None,:]
            next_state = agent.poly.fit_transform(next_state.reshape(1, 4))
            grad = agent.grad(probs, action, state)

            grads.append(grad)
            rewards.append(reward)

            score += reward
            state = next_state

            if done:
                break

        agent.update_with(grads, rewards)
        complete_scores.append(score)

    env.close()
    plt.plot(np.arange(NUM_EPISODES),
             complete_scores)
    plt.savefig('image1.png')


if __name__ == '__main__':
    main(None)

.

.

问题

我正在尝试做几乎相同的示例，但使用 Sigmoid 激活（只是为了简单起见）。 这就是我需要做的。将模型中的激活从 softmax 切换到 sigmoid。 这肯定会起作用（根据下面的解释）。但是我的 Policy Gradient 模型什么也没学到，而且保持随机。有什么建议吗？

import gym
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import PolynomialFeatures

NUM_EPISODES = 4000
LEARNING_RATE = 0.000025
GAMMA = 0.99


# noinspection PyMethodMayBeStatic
class Agent:
    def __init__(self):
        self.poly = PolynomialFeatures(1)
        self.w = np.random.rand(5, 1) - 0.5

    # Our policy that maps state to action parameterized by w
    # noinspection PyShadowingNames
    def policy(self, state):
        z = np.sum(state.dot(self.w))
        return self.sigmoid(z)

    def sigmoid(self, x):
        s = 1 / (1 + np.exp(-x))
        return s

    def sigmoid_grad(self, sig_x):
        return sig_x * (1 - sig_x)

    def grad(self, probs, action, state):
        dsoftmax = self.sigmoid_grad(probs)
        dlog = dsoftmax / probs
        grad = state.T.dot(dlog)
        grad = grad.reshape(5, 1)
        return grad

    def update_with(self, grads, rewards):
        if len(grads) < 50:
            return
        for i in range(len(grads)):
            # Loop through everything that happened in the episode
            # and update towards the log policy gradient times **FUTURE** reward

            total_grad_effect = 0
            for t, r in enumerate(rewards[i:]):
                total_grad_effect += r * (GAMMA ** r)
            self.w += LEARNING_RATE * grads[i] * total_grad_effect


def main(argv):
    env = gym.make('CartPole-v0')
    np.random.seed(1)

    agent = Agent()
    complete_scores = []

    for e in range(NUM_EPISODES):
        state = env.reset()[None, :]
        state = agent.poly.fit_transform(state)

        rewards = []
        grads = []
        score = 0

        while True:

            probs = agent.policy(state)
            action_space = env.action_space.n
            action = np.random.choice(action_space, p=[1 - probs, probs])

            next_state, reward, done, _ = env.step(action)
            next_state = next_state[None, :]
            next_state = agent.poly.fit_transform(next_state.reshape(1, 4))

            grad = agent.grad(probs, action, state)
            grads.append(grad)
            rewards.append(reward)

            score += reward
            state = next_state

            if done:
                break

        agent.update_with(grads, rewards)
        complete_scores.append(score)

    env.close()
    plt.plot(np.arange(NUM_EPISODES),
             complete_scores)
    plt.savefig('image1.png')


if __name__ == '__main__':
    main(None)

绘制所有学习保持随机。调整超参数没有任何帮助。在示例图片下方。

参考文献：

1. Deep Reinforcement Learning: Pong from Pixels

1. An introduction to Policy Gradients with Cartpole and Doom

1. Deriving Policy Gradients and Implementing REINFORCE

4. Machine Learning Trick of the Day (5): Log Derivative Trick 12

---

更新

似乎下面的答案可以从图形中做一些工作。但这不是对数概率，甚至不是策略的梯度。并改变了 RL 梯度策略的整个目的。请检查上面的参考资料。我们下图后声明。

我需要我的策略的日志函数的梯度（这只是权重和sigmoid 激活）。如有任何问题，请告诉我。

Answer 1

问题在于grad 方法。

def grad(self, probs, action, state):
    dsoftmax = self.sigmoid_grad(probs)
    dlog = dsoftmax / probs
    grad = state.T.dot(dlog)
    grad = grad.reshape(5, 1)
    return grad

在原始代码中，Softmax 与 CrossEntropy 损失函数一起使用。当您将激活切换到 Sigmoid 时，正确的损失函数变为 Binary CrossEntropy。现在，grad 方法的目的是计算损失函数 wrt 的梯度。权重。 Sparing 细节，适当的渐变由(probs - action) * state 在您的程序术语中给出。最后一件事是添加减号——我们想要最大化损失函数的负数。

正确的grad 方法如下：

def grad(self, probs, action, state):
    grad = state.T.dot(probs - action)
    return -grad

您可能想要添加的另一个更改是提高学习率。 LEARNING_RATE = 0.0001 和 NUM_EPISODES = 5000 将产生以下情节：

如果权重使用零均值和小方差的高斯分布初始化，收敛速度会快得多：

def __init__(self):
    self.poly = PolynomialFeatures(1)
    self.w = np.random.randn(5, 1) * 0.01

更新

添加了完整的代码来重现结果：

import gym
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import PolynomialFeatures

NUM_EPISODES = 5000
LEARNING_RATE = 0.0001
GAMMA = 0.99


# noinspection PyMethodMayBeStatic
class Agent:
    def __init__(self):
        self.poly = PolynomialFeatures(1)
        self.w = np.random.randn(5, 1) * 0.01

    # Our policy that maps state to action parameterized by w
    # noinspection PyShadowingNames
    def policy(self, state):
        z = np.sum(state.dot(self.w))
        return self.sigmoid(z)

    def sigmoid(self, x):
        s = 1 / (1 + np.exp(-x))
        return s

    def sigmoid_grad(self, sig_x):
        return sig_x * (1 - sig_x)

    def grad(self, probs, action, state):
        grad = state.T.dot(probs - action)
        return -grad

    def update_with(self, grads, rewards):
        if len(grads) < 50:
            return
        for i in range(len(grads)):
            # Loop through everything that happened in the episode
            # and update towards the log policy gradient times **FUTURE** reward

            total_grad_effect = 0
            for t, r in enumerate(rewards[i:]):
                total_grad_effect += r * (GAMMA ** r)
            self.w += LEARNING_RATE * grads[i] * total_grad_effect


def main(argv):
    env = gym.make('CartPole-v0')
    np.random.seed(1)

    agent = Agent()
    complete_scores = []

    for e in range(NUM_EPISODES):
        state = env.reset()[None, :]
        state = agent.poly.fit_transform(state)

        rewards = []
        grads = []
        score = 0

        while True:

            probs = agent.policy(state)
            action_space = env.action_space.n
            action = np.random.choice(action_space, p=[1 - probs, probs])

            next_state, reward, done, _ = env.step(action)
            next_state = next_state[None, :]
            next_state = agent.poly.fit_transform(next_state.reshape(1, 4))

            grad = agent.grad(probs, action, state)
            grads.append(grad)
            rewards.append(reward)

            score += reward
            state = next_state

            if done:
                break

        agent.update_with(grads, rewards)
        complete_scores.append(score)

    env.close()
    plt.plot(np.arange(NUM_EPISODES),
             complete_scores)
    plt.savefig('image1.png')


if __name__ == '__main__':
    main(None)