GridWorld

The code in this section is adapted from https://github.com/sachag678.

We begin by demonstrating the basic TF-Agents functionality in the GridWorld environment. RL problems are best studied in the context of either games (where we have a clearly defined set of rules and fully observable context), or toy problems such as GridWorld. Once the basic concepts are clearly defined in a simplified but non-straightforward environment, we can move to progressively more challenging situations.

The first step is to define a GridWorld environment: this is a 6x6 square board, where the agent starts at (0,0), the finish is at (5,5), and the goal of the agent is to find the path from the start to the finish. Possible actions are moves up/down/left/right. If the agent lands on the finish, it receives a reward of 100, and the game terminates after 100 steps if the end was not reached by the agent. An example of the GridWorld "map" is provided here:

Figure...