Abstract

Abstract: Automating the development of machine learning algorithms (i.e., meta-learning) has the potential to unlock new frontiers in the field. However, our ability to learn to discover has been limited by a focus on small, static benchmarks. Motivated by how procedural generation unlocked generalist agents in reinforcement learning, this talk will explore how a similar approach can be applied to algorithm discovery in machine learning. Specifically, I will introduce DiscoGen, a new procedural generator of algorithm discovery tasks. Using DiscoGen, we demonstrate how agents used for algorithm discovery can themselves be optimised in a meta-meta-loop. DiscoGen further establishes principled task design for the field, in particular emphasising the need for meta-train and meta-test distinctions. Finally, the talk will discuss future research ideas enabled by DiscoGen, such as training algorithm world models or other means for optimising discovery agents.

Bio: Alexander D. Goldie is a final-year PhD student at the University of Oxford, co-supervised by Jakob N. Foerster (FLAIR) and Shimon Whiteson (WhiRL). His research focuses on the automated discovery of machine learning algorithms (meta-learning), using black-box evolution, symbolic evolution, or agentic LLMs. More recently, he has explored how agents can themselves be optimised in a meta-meta-learning process.