强化学习中子进程调用atari游戏是否受父进程中设置的随机种子影响

python中numpy.random.seed设置随机种子是否影响子进程

============================================

代码：

from ale_python_interface import ALEInterface
import numpy as np
np.random.seed(1111)
import cv2
import time

filename = "atari_roms" + "/" + "pong" + ".bin"

ale_int = ALEInterface()
ale_int.setInt(b"random_seed", 1111)
ale_int.setFloat(b"repeat_action_probability", 0.0)
ale_int.setInt(b"frame_skip", 1)
ale_int.setBool(b"color_averaging", False)
ale_int.loadROM(str.encode(filename))
num_actions = len(ale_int.getMinimalActionSet())
legal_actions = ale_int.getMinimalActionSet()
h, w = ale_int.getScreenDims()
gray_screen = np.zeros((h, w, 1), dtype=np.uint8)


ale_int.reset_game()

pre_screen = None
for i in range(30):
    ale_int.act(legal_actions[0])
    ale_int.getScreenGrayscale(gray_screen)

for i in range(1000):
    pre_screen = np.copy(gray_screen)
    ale_int.act(legal_actions[np.random.randint(len(legal_actions))])
    ale_int.getScreenGrayscale(gray_screen)
    # cv2.imshow("Example Image", gray_screen)
    # time.sleep(0.01)

    # print(gray_screen)
    det = np.sum(pre_screen - gray_screen)
    print(det, ale_int.game_over())

运行结果：

=============================

对比代码：

from multiprocessing import Process

from ale_python_interface import ALEInterface
import numpy as np
np.random.seed(1111)
# import cv2
# import time

filename = "atari_roms" + "/" + "pong" + ".bin"

ale_int = ALEInterface()
ale_int.setInt(b"random_seed", 1111)
ale_int.setFloat(b"repeat_action_probability", 0.0)
ale_int.setInt(b"frame_skip", 1)
ale_int.setBool(b"color_averaging", False)
ale_int.loadROM(str.encode(filename))
num_actions = len(ale_int.getMinimalActionSet())
legal_actions = ale_int.getMinimalActionSet()
h, w = ale_int.getScreenDims()
gray_screen = np.zeros((h, w, 1), dtype=np.uint8)

ale_int.reset_game()


class NN(Process):
    def __init__(self, id, ale):
        super(NN, self).__init__()
        self.id = id
        self.ale = ale
    def run(self):
        super(NN, self).run()

        ale_int = self.ale


        num_actions = len(ale_int.getMinimalActionSet())
        legal_actions = ale_int.getMinimalActionSet()
        h, w = ale_int.getScreenDims()
        gray_screen = np.zeros((h, w, 1), dtype=np.uint8)

        pre_screen = None
        for i in range(30):
            ale_int.act(legal_actions[0])
            ale_int.getScreenGrayscale(gray_screen)

        for i in range(1000):
            pre_screen = np.copy(gray_screen)
            ale_int.act(legal_actions[np.random.randint(len(legal_actions))])
            ale_int.getScreenGrayscale(gray_screen)
            # cv2.imshow("Example Image", gray_screen)
            # time.sleep(0.01)

            # print(gray_screen)
            det = np.sum(pre_screen - gray_screen)
            print(det, ale_int.game_over())


ps = [NN(i, ale_int) for i in range(1)]
for p in ps:
    p.start()

for p in ps:
    p.join()

运行结果：

PS:

可以看到，在python中子进程生成时会copy父进程中的对象，哪怕是atari游戏这种调用C语言扩展模块的对象也会被copy状态给子进程，这个和其他python中对象一样；这个特点和python中numpy.random.seed设置随机种子是否影响子进程相一致。

注意，上面代码中设置numpy和atari游戏的随机种子状态的代码为：

PS:

扩展一下，正因为python在生成子进程时会copy父进程状态这一特点，所以在生成子进程时我们需要在子进程中设置随机种子，并且在子进程中设置随机种子时要保证各个子进程被传入一个不同的数值；要注意在各个子进程中time.time()的数值也都是相同的，如果不能从父进程为各个子进程传入一个不同的数值，那么各个进程运行起来所使用的随机种子会是一致的，这样会影响最终的运算结果。给出具体例子：