Hands-On Neuroevolution with Python

Creating the network graph

The TensorFlow network graph is created by the make_net function, which receives the ANN model constructor, the GPU device identifier, and the batch size as input parameters. The network graph is created as follows:

1. We'll start by creating the controller ANN model and the game evaluation environment:

    self.model = model_constructor()    …    with tf.variable_scope(None, default_name='model'):        with tf.device(‘/cpu:0'):            self.env = self.make_env_f(self.batch_size)

2. Next, we'll create placeholders so that we can receive values during network graph evaluation. Also, we'll create an operator to reset the game before the start of the new game episode:

        self.placeholder_indices = tf.placeholder(tf.int32,                                                     shape=(None, ))