In this talk, I will present the framework of online convex reinforcement learning. After introducing the framework and

discussing how it differs from standard reinforcement learning, I will present an algorithm based on Online Mirror Descent

over state-action distributions, together with its theoretical guarantees. I will then present an application to demand-side

management, which aims to adapt the energy consumption of devices to better match the production of renewable

energy. Finally, I will briefly outline several extensions, including non-stationary environments, non-episodic settings, and

multinomial logistic models for the transition kernel.

Refs : https://arxiv.org/pdf/2505.07303

          https://link.springer.com/article/10.1007/s10957-025-02658-9