eblancoh

eblancoh

eblancoh

eblancoh

eblancoh

eblancoh

class ReplayMemory:

    ...

    def update_all_q_values(self):
        """
        Update all Q-values in the replay-memory.

        When states and Q-values are added to the replay-memory, the
        Q-values have been estimated by the Neural Network. But we now

eblancoh

class NeuralNetwork:

    def __init__(self):

        # Initializer for the layers in the Neural Network.
        # If you change the architecture of the network, particularly
        # if you add or remove layers, then you may have to change
        # the stddev-parameter here. The initial weights must result
        # in the Neural Network outputting Q-values that are very close
        # to zero - but the network weights must not be too low either

eblancoh

class NeuralNetwork:

    def __init__(self):

        ...

        # TensorFlow has a built-in loss-function for doing regression:
        # self.loss = tf.nn.l2_loss(self.q_values - self.q_values_new)
        # But it uses tf.reduce_sum() rather than tf.reduce_mean()
        # which is used by PrettyTensor. This means the scale of the

eblancoh

def _rgb_to_black_and_white(image):
    """
    Convert an RGB-image into gray-scale using a formula from Wikipedia:
    https://en.wikipedia.org/wiki/Grayscale
    """

    # Get the separate colour-channels.
    r, g, b = image[:, :, 0], image[:, :, 1], image[:, :, 2]

    # Convert to gray-scale using the Wikipedia formula.

eblancoh

class EpsilonGreedy:
    """
    The epsilon-greedy policy either takes a random action with
    probability epsilon, or it takes the action for the highest
    Q-value.

    If epsilon is 1.0 then the actions are always random.
    If epsilon is 0.0 then the actions are always argmax for the Q-values.

    Epsilon is typically decreased linearly from 1.0 to 0.1 during training