Training the Controller

How do we use this reward signal to update the Controller? Remember, this reward signal is not differentiable like a loss function in supervised learning; we may not backpropagate this through the Controller on its own. Instead, we employ a policy gradient method called REINFORCE to iteratively update the Controller parameters, . In REINFORCE, the gradient of the reward function, J, with respect to the parameters of the Controller, , is defined as follows:

You may recall seeing a similar expression in Chapter 6, Learning ...

Get Python Reinforcement Learning Projects now with the O’Reilly learning platform.

O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.