Code for the tutorial on medium: Reinforcement learning on Reversing Stones

initialise.py

inputs_units = BOARD_SIZE[0] * BOARD_SIZE[1]
hidden_units = BOARD_SIZE[0] * BOARD_SIZE[1]
output_units = BOARD_SIZE[0] * BOARD_SIZE[1]

def initialise_tf():
    global input_positions, labels, learning_rate, W1, b1, h1, W2, b2, logits, probabilities, cross_entropy, train_step
    input_positions = tf.placeholder(tf.float32, shape=(1, inputs_units))
    labels          = tf.placeholder(tf.int64)
    learning_rate   = tf.placeholder(tf.float32, shape=[])

    # Generate hidden layer
    W1 = tf.Variable(tf.truncated_normal([inputs_units, hidden_units], stddev=0.1 / inputs_units**0.5))
    b1 = tf.Variable(tf.zeros([1, hidden_units]))
    h1 = tf.tanh(tf.matmul(input_positions, W1) + b1)

    # Second ## -- linear classifier for action logits
    W2 = tf.Variable(tf.truncated_normal([hidden_units, output_units], stddev=0.1 / hidden_units**0.5))
    b2 = tf.Variable(tf.zeros([1, output_units]))

    logits = tf.matmul(h1, W2) + b2
    probabilities = tf.nn.softmax(logits)

    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels, name='xentropy')
    train_step = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cross_entropy)