Being the final chapter of this book, this chapter provided summaries of key learning algorithms that are currently state of the art in this domain. We looked at the core concepts behind three different state-of-the-art algorithms, each with their own unique elements and their own categories (actor-critic/policy based/value-function based).
Specifically, we discussed the deep deterministic policy gradient algorithm, which is an actor-critic architecture method that uses a deterministic policy rather than the usual stochastic policy, and achieves good performance on several continuous control tasks.
We then looked at the PPO algorithm, which is a policy gradient-based method that uses a clipped version of the TRPO objective and learns a monotonically better and stable policy, and has been successfully used in very high-dimensional environments such as DOTA II.
Finally...