Simple clicking approach
As the first demo, let's implement a simple Asynchronous Advantage Actor-Critic (A3C) agent, which decides where it should click on given the image observation. This approach can solve only a small subset of the full MiniWoB suite and we'll discuss restrictions of this approach later. For now, it will allow us to get a better understanding of the problem.
As with the previous chapter, due to size of the code, I won't put a complete source code here. We'll focus on the most important functions and give the rest as an overview. The complete source code is available in the GitHub repository https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On.
Grid actions
When we talked about OpenAI Universe's architecture and organization, it was mentioned that the richness and flexibility of the action space creates lots of challenges for the RL agent. MiniWoB's active area inside the browser is just 160x210Â (exactly the same dimension...