Reinforcement learning for Control

Stochastic Policy Search

Policy network

In the same way as we used data-driven optimization to tune the gains $K_P,K_I,K_D$ in the PID controllers, we can use the same approach to tune (or train) the parameters (or weights) $\boldsymbol{\theta}$ of a neural network.

In the first part a relatively simple neural network controller in PyTorch is built.

Reinforce algorithm

In order to implement and apply the Reinforce algorithm, the following steps are performed:

• Create a policy network that uses transfer learning
• Create an auxiliary function that selects control actions out of the distribution
• Create an auxilary function that runs multiple episodes per epoch
• Finally, put all the pieces together into a function that computes the Reinforce algorithm