The simple clicking approach
To get started with web navigation, let’s implement a simple A3C agent that decides where it should click given the image observation. This approach can solve only a small subset of the full MiniWoB suite, and we will discuss the restrictions of this approach later. For now, it will allow us to get a better understanding of the problem.
As with the previous chapter, I won’t discuss the complete source code here. Instead, we will focus on the most important functions and I will provide a brief overview of the rest. The complete source code is available in the GitHub repository.
Grid actions
When we talked about MiniWoB architecture and organization, we mentioned that the richness and flexibility of the action space creates a lot of challenges for the RL agent. The active area inside the browser is just 210 × 160 pixels, but even with such a small area, our agent could be asked to move the mouse, perform...
