Summary

Extending from where we left off with DQN, we looked at ways of extending this model with CNN and adding additional networks to create double DQN and dueling DQN, or DDQN. Before exploring CNN, we looked at what visual observation encoding is and why we need it. Then, we briefly introduced CNN and used the TensorSpace Playground to explore some well-known, state-of-the-art models. Next, we added CNN to a DQN model and used that to play the Atari game environment Pong. After, we took a closer look at how we could extend DQN by adding another network as the target and adding another network to duel against or to contradict the other network, also known as the dueling DQN or DDQN. This introduced the concept of advantage in choosing an action. Finally, we looked at extending the experience replay buffer so that we can prioritize events that get captured there. Using this...