some changes/ additions to ch 35 (RL)

[murphyk]

Sec 35.3.5 (DDPG)

• Currently the text just says "The DDPG algorithm of [Lil+16], which stands for “deep deterministic policy gradient”, uses the DQN method (Section 35.2.6) to update Q that is represented by deep neural networks. " We expand on this a little bit, since DDPG is quite widely used.

Sec 35.4 (MBRL)

• Significant expansion and restructuring of the content

• Sec 35.5.3 (deadly triad)

• Added brief discussion of gradient TD and target networks to stabilize off-policy learning.

Sec 35.5.4 (new: offpolicy in practice)

• add very short new section listing common off-policy methods, for ease of reference.

Sec 35.6 (control as inference)

• Added clearer subsection headings, to add more structured
• Moved the subsection on 'imitation learning' into its own sec 35.7.

Sec 35.7 (new: imitation learning)

• This now contains the content that used to be in 35.6.3

Sec 35.8 (new: "Other topics in RL")

• Added brief discussions of various topics, such as GVF, temporal abstraction (options), partial observability, reward functions (including shaping and hacking), and offline RL.

[murphyk]

To avoid too much divergence from the original text, I have rolled back these changes. This new content will be added to a new tutorial on RL that I am writing.