Policy gradient methods for ordinal policies

Simón Weinberger, Jairo Cugliari

In reinforcement learning, the softmax parametrization is the standard approach for policies over discrete action spaces. However, it fails to capture the order relationship betwee…