Memahami Kebijakan Gradien Berasal

Saya mencoba membuat ulang contoh Gradient Kebijakan yang sangat sederhana, dari sumbernya, blog Andrej Karpathy . Dalam artikel itu, Anda akan menemukan contoh dengan CartPole dan Policy Gradient dengan daftar bobot dan aktivasi Softmax. Ini adalah contoh gradien kebijakan CartPole yang saya buat dan sangat sederhana, yang berfungsi sempurna .

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)

.

.

Pertanyaan

Saya coba lakukan, contoh yang hampir sama tetapi dengan aktivasi Sigmoid (hanya untuk kesederhanaan). Hanya itu yang perlu saya lakukan. Beralih aktivasi dalam model dari softmaxke sigmoid. Yang pasti berhasil (berdasarkan penjelasan di bawah). Tetapi model Gradient Kebijakan saya tidak mempelajari apa pun dan tetap acak. Ada saran?

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)

Merencanakan semua pembelajaran dilakukan secara acak. Tidak ada yang membantu dengan menyetel parameter hiper. Di bawah gambar contoh.

Referensi :

1) Pembelajaran Penguatan Dalam: Pong from Pixels

2) Pengantar Gradien Kebijakan dengan Cartpole dan Doom

3) Turunkan Gradien Kebijakan dan Implementasi REINFORCE

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

MEMPERBARUI

Sepertinya jawaban di bawah ini dapat melakukan beberapa pekerjaan dari grafik. Tapi itu bukan probabilitas log, dan bahkan tidak gradien kebijakan. Dan mengubah seluruh tujuan Kebijakan Gradien RL. Silakan periksa referensi di atas. Setelah gambar kita pernyataan selanjutnya.

Saya perlu mengambil Gradien fungsi Log Kebijakan saya (yang hanya berupa bobot dan sigmoidaktivasi).

Masalahnya dengan gradmetode.

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

Dalam kode asli Softmax digunakan bersama dengan fungsi kerugian CrossEntropy. Ketika Anda beralih aktivasi ke Sigmoid, fungsi kerugian yang tepat menjadi Binary CrossEntropy. Sekarang, tujuan dari gradmetode ini adalah untuk menghitung gradien dari fungsi kehilangan wrt. beban. Menyelamatkan detail, gradien yang tepat diberikan oleh (probs - action) * statedalam terminologi program Anda. Hal terakhir adalah menambahkan tanda minus - kami ingin memaksimalkan negatif dari fungsi kerugian.

gradMetode yang tepat sebagai berikut:

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

Perubahan lain yang mungkin ingin Anda tambahkan adalah untuk meningkatkan tingkat belajar. LEARNING_RATE = 0.0001dan NUM_EPISODES = 5000akan menghasilkan plot berikut:

Konvergensi akan jauh lebih cepat jika bobot diinisialisasi menggunakan distribusi Gaussian dengan mean nol dan varians kecil:

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

MEMPERBARUI

Menambahkan kode lengkap untuk mereproduksi hasil:

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)