Global Optimality Guarantees For Policy Gradient Methods

Jalaj Bhandari, Daniel Russo

Policy gradients methods apply to complex, poorly understood, control problems by performing stochastic gradient descent over a parameterized class of polices. Unfortunately, even…