As always, try and complete a minimum of two to three of these exercises on your own, and for your own benefit. While this is a hands-on book, it always helps to spend a little more time applying your knowledge to new problems.

Complete the following exercises on your own:

1. Go through and explore the VisualPushBlock example. This example is quite similar to the Hallway, and is a good analog to play with.
2. Modify the Hallway example's HallwayAgent script to use more scanning angles, and thus more vector observations.
3. Modify the Hallway example to use a combined sensor sweep and visual observation input. This will require you to modify the learning brain configuration by adding a camera, and possibly updating some hyperparameters.
4. Modify other visual observation environments to use some form of vector observation. A good example to try this on is the VisualPushBlock example.
5. Modify the visual observation camera space to be larger or smaller than 84 x 84 pixels, and to use, or not use, gray scaling. This is a good exercise to play with when testing more complex or simpler CNN network architectures.
6. Modify the create_visual_observation_encoder convolutional encoding function so that it can use different CNN architectures. These architectures may be as simple or complex as you want.
7. Modify the create_visual_observation_encoder convolutional encoding function to use different levels and amounts of pooling layers. Try and use pooling after every convolutional layer to explore its effect on training.
8. Disable and enable recurrent networks on one or two of the other example's environments and explore the effect this has.
9. Play with the sequence_length and memory_size parameters with recurrent enabled to see the effect that different sequence lengths have on agent performance. Be sure to increase the memory_size parameter if you increase the sequence_length.
10. Consider adding additional vector or visual observations to the agent. After all, an agent doesn't have to have only a single form of sensory input. An agent could always detect the direction it is in, or perhaps it may have other forms of sensory input, such as being able to listen. We will give an agent the ability to listen in a later chapter, but try and implement this yourself.

Remember, these exercises are provided for your benefit and enjoyment, so be sure to try at least a couple.