Joost Huizinga
Uber AI (United States), University of Wyoming, OpenAI (United States)
Papers
8
Total Citations
687
H-Index
7
About
Joost Huizinga is a leading researcher at the intersection of reinforcement learning (RL), evolutionary computation, and artificial intelligence, whose work has fundamentally reshaped how we approach hard-exploration problems. He is best known as the co-creator of the **Go-Explore** algorithm, a paradigm-shifting approach that tackles environments with sparse or deceptive rewards—such as the notoriously difficult Atari games Montezuma’s Revenge and Pitfall—by first returning to promising states before exploring. The original Go-Explore paper (2019) has garnered over 228 citations, while its refined successor, “First return, then explore,” has accumulated over 212 citations, cementing Huizinga’s impact on the RL community. Beyond exploration, he has made seminal contributions to understanding **modularity, regularity, and hierarchy in evolved neural networks**, demonstrating how structural organization can improve evolutionary optimization (123 citations for his work on the evolutionary origins of hierarchy). Huizinga also contributed to **Video PreTraining (VPT)**, a method that learns to act by watching unlabeled online videos (50+ citations), and developed the **Combinatorial Multiobjective Evolutionary Algorithm** for evolving multimodal robot behavior. His research elegantly bridges evolutionary biology and machine learning, offering practical algorithms that enable agents to discover complex behaviors autonomously.
Research Focus
Key Achievements
Top Papers
- 1Go-Explore: a New Approach for Hard-Exploration Problems228 citations · 2019
- 2First return, then explore212 citations
- 3The Evolutionary Origins of Hierarchy123 citations · 2016
- 4Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos50 citations · 2022
- 5Evolving neural networks that are both modular and regular43 citations · 2014
- 6
- 7
- 8