Python Deep Learning - Second Edition

Policy improvements

Once we're able to evaluate a policy, let's look at how to improve it. This task is also known as control. We'll assume that the policy is represented as a table, where the best actions are stored for each state (tabular solution). We'll also assume that we have an already-existing value function,(the step described in the preceding section), and a policy, π. For each state, s, we'll do the following:

1. Assume that we take all possible actions starting from s. That also includes the action selected by the policy. Using the action-value Bellman equation, for each action we'll compute the expected returns if we take that action ...