Hands-On Machine Learning for Algorithmic Trading

The DQN architecture

DQN was first applied to the Atari domain with high-dimensional image observations and relied on convolutional layers. The LL's lower-dimensional state representation makes fully-connected layers a better choice (see Chapter 16, Deep Learning).

More specifically, the network maps eight inputs to four outputs that correspond to the Q values for each action, so that it only takes a single forward pass to compute the action values. The DQN is trained on the preceding loss function using the Adam optimizer. The agent's DQN uses 3 densely connected layers with 256 units each and L2 activity regularization. Using a GPU using the TensorFlow Docker image (see Chapter 16, Deep Learning, and Chapter 18, Recurrent Neural Networks ...