Beyond Trial & Error: A Tutorial on Automated Reinforcement Learning

This page serves as an overview of André Biedenkapp and Theresa Eimer’s tutorial on AutoRL at ECAI 2024. We will add slides and code examples at a later point.

Outline

  1. Introduction and algorithmic part on AutoRL
    • Motivation: Why does AutoRL matter?
    • Formal definition of AutoRL
    • Categories of AutoRL approaches (e.g. learning to learn, environment design, etc.)
    • Properties of AutoRL landscapes
    • What are AutoRL-specific challenges compared to AutoML for supervised learning?
    • Why are dynamic configuration approaches important for RL, and how do we learn them?
  2. Practical guidelines and case study of hyperparameters:
    • Examples of successful AutoRL, DAC and online approaches
    • Evaluation and Generalization of AutoRL
    • HPO for RL
    • Hyperparameters and experimental design
    • Forms of optimization with pros and cons (AC methods, PBT, heuristics, meta-gradients, etc.)
    • Combining HPO with other AutoRL domains and why this is important for RL generally

Speakers

Theresa Eimer is a Ph.D. student at the Leibniz University of Hannover focusing on the intersection of AutoML and Reinforcement Learning. Her goal is to make Reinforcement Learning work out of the box through better generalization and automatic configuration.

André Biedenkapp is a Postdoctoral Fellow at the University of Freiburg focusing on dynamic configuration with and for deep reinforcement learning. With his work, he aims to democratize deep reinforcement learning.