Policy Gradient methods in RL optimize a parameterized policy directly by ascending the gradient of expected rewards. This is different from value-based methods that learn a value function. Policy gradient algorithms can handle continuous action spaces and learn stochastic policies.