Deep Learning with Keras

The deep Q-network as a Q-function

We know that our Q-function is going to be a neural network, the natural question is: what kind? For our simple example game, each state is represented by four consecutive black and white screen images of size (80, 80), so the total number of possible states (and the number of rows of our Q-table) is 2^80x80x4. Fortunately, many of these states represent impossible or highly improbable pixel combinations. Since convolutional neural networks have local connectivity (that is, each neuron is connected to only a local region of its input), it avoids these impossible or improbable pixel combinations. In addition, neural networks are generally very good at coming up with good features for structured data such as ...