Tutorial

Description

Learning agents, i.e., software systems based on Deep Reinforcement Learning, have already firmly established themselves in a multitude of applications for smart grids. They promise to provide resilient strategies for the operation of power grids, being able to adapt themselves and react to events that were not foreseen by their creators. DRL agents need to be trained in simulations; the more extensive, the better. In order to learn faster and to avoid the sampling bias problem, autocurriculum setups, i.e., agents training against each other, have been introduced.
This tutorial will provide a practical introduction to deep reinforcement learning with a special focus on autocurriculum setups. It will showcase an experiment considering voltage control in a realistic grid, introducing a comprehensive simulation and experimentation framework as one way to approach the topic.

Curriculum Vitae

Eric is computer scientist by heart. He obtained his diploma in 2010; his studies initially focused on computer and communication networks. During is PhD, he devised a Multi-Agent System for guaranteed optimal real power equilibria in distribution grids with a high amount of volatile, distributed, renewable energy resources.
In 2017, after his PhD, he joined the computer science institute OFFIS in Oldenburg, Germany. Here, he created Adversarial Resilience Learning (ARL), a methodology based on Deep Reinforcement Learning to analyize a cyber-physical energy system for weaknesses and let the agents reliably learn strategies for a resilient operation.
Since 2022, Eric leads his own junior research group at the University of Oldenburg, Germany, focusing on extended agent architecture based on ARL that allow for guarantees and explainability and supervising PhD candidates in this particular area of research. Eric is member of Germany's federal platform for artificial intelligence (PLS), of the Transatlantic Cyber Forum, the German standardization organization (DIN), and serves in the TCP of IARIA's ENERGY conference series, as well as the ACM e-Energy.