Reinforcement Learning Algorithms with Python

The alternative

The answer to this question is no.

As we learned in Chapter 10, Imitation Learning with the DAgger Algorithm, by reducing policy learning to an imitation problem using backpropagation and SGD, we can learn about a discriminative model from an expert in order to predict which actions to take next. Still, this involves backpropagation and requires an expert that may not always be available.

Another general subset of algorithms for global optimization does exist. They are called EAs, and they aren't based on backpropagation and don't require any of the other two principles, namely temporal credit assignment and noisy actions. Furthermore, as we said in the introduction to this chapter, these evolutionary algorithms are very ...