The exercises in this chapter are a mix of working with ML-Agents and building your own testing analysis platform. As such, choose one or two exercises that make sense for you to complete on your own from the following list:

1. Configure the TestingAgent to use a different camera for its visual observation input.
2. Enable Curiosity Learning on the agent's brain.
3. Set up the TestingAgent to control a different vehicle.

4. Set up the TestingAgent to run on another vehicle and let ML-Agents control both of the agents simultaneously.
5. Add additional tracking analytics custom events for the agents. Perhaps track the distance that the agent travels versus its lifetime. This will provide a speed factor that can also denote the agent's efficiency. An agent that hits a goal quicker will have a better speed factor.
6. Enable online imitation learning by adding a second vehicle with a learning agent. If you need to, go back and review the setup of the tennis scene.
7. Set up the Academy to use curriculum learning. Perhaps allow the virtual goal deployment box to grow in size over training iterations (by 10%, or some other factor). This will allow the goals to disperse farther and make it more difficult for the agent to find.
8. Modify the visual observation input that the brains are using to 184 x 184, the new standard, and see what effect this has on agent training.
9. Modify the visual observation convolutional encoding network, as we did in Chapter 7, Agents and the Environment, to use more layers and/or different filtering.
10. Apply this testing framework to your own game. Be sure to also add the analytics, so that you can track training and player usage.

These exercises are more involved than those in the previous chapters, since this is a big and important chapter. In the next section, we will review what you learned and covered in this chapter.