Hands-On Neuroevolution with Python

The CNN architecture to train the Atari playing agent

In our experiment, we'll use a CNN architecture consisting of three convolutional layers with 32, 64, and 64 channels, followed by a fully connected layer with 512 units and the output layer with the number of layers corresponding to the number of game actions. The convolutional layers have 8 x 8, 4 x 4, and 3 x 3 kernel sizes and use strides of 4, 2, and 1, respectively. The ReLU non-linearity follows all the convolutional and fully connected layers.

The source code to create the described network graph model using the TensorFlow framework is defined as follows:

class LargeModel(Model):    def _make_net(self, x, num_actions):        x = self.nonlin(self.conv(x, name='conv1', num_outputs=32,  kernel_size=8, ...