Hands-On Reinforcement Learning for Games

Using PPO with recurrent networks

In Chapter 7, Going Deeper with DDQN, we saw how we could interpret visual state using a concept called convolutional neural networks (CNNs). CNN networks are used to detect features in visual environments such as Atari games. While this technique allowed us to play any of a number of games with the same agent, the added CNN layers took much more time to train. In the end, the extra training time wasn't worth the cool factor of playing Atari games. However, there are other network structures we can put on top of our networks in order to make better interpretations of state. One such network structure is called recurrent networks. Recurrent network layers allow us to add the concept of context or time in our ...