Policy Gradient Methods
Every algorithm so far learns a value function first, and derives a policy from it only indirectly (act greedily with respect to the values). Policy gradient methods skip the middleman entirely — parameterise the policy directly, and take gradients of expected return with respect to those parameters.