This thread confuses problems and algorithms. Based on its name, "Reinforcement Learning" is learning from (possibly-delayed) bandit feedback. That is, you only obtain feedback on the actions you take, and the feedback does not tell you what the correct action should be

ReLu?

For everything that could be simulated there's RL

Probably just REINFORCE, which is not strictly for reinforcement learning but also used for non-differentiable objective functions.

When you only have a hammer...

I’ve always thought of RL (where the difference between estimated and observed returns is calculated) as performing gradient descent (even in straight MDP). SGD for supervised problems is a special case of RL backup rules (not the other way around).

Doesnt RL still use backpropagation though?