Joel Lehman is a computer scientist and artificial intelligence researcher affiliated with OpenAI, where he works on advancing machine learning and artificial intelligence. He is best known for co-creating novelty search and POET, two influential algorithms that explore open-endedness in AI systems. His research focuses on how simple evolutionary principles can generate complex behaviors and solutions without explicit objective functions.
Lehman's work sits at the intersection of artificial intelligence, machine learning, and neural networks. He has contributed to both foundational theory and practical applications, including large language models and reinforcement learning. His ideas have shaped how researchers think about exploration, diversity, and the potential for AI to innovate beyond predefined goals.
Early Life and Education
Joel Lehman received his PhD in computer science from the University of Texas at Austin, where he worked under the supervision of Risto Miikkulainen. His doctoral research focused on evolutionary computation and the use of novelty search to avoid deceptive fitness landscapes. He also holds a bachelor's degree in computer science from Carnegie Mellon University.
During his graduate studies, Lehman developed a deep interest in open-ended evolution, a concept that draws from biological evolution to create systems that continuously generate new and increasingly complex behaviors. This interest would later define much of his career.
Novelty Search
Novelty search is an algorithm that rewards agents for producing behaviors that are novel or different from those seen before, rather than for achieving a specific goal. Lehman co-created this approach with Kenneth O. Stanley in the mid-2000s. The key insight is that in many complex problems, the path to a solution is not obvious, and optimizing for novelty can lead to discoveries that would be missed by traditional objective-based search.
In a series of papers starting around 2008, Lehman and Stanley demonstrated that novelty search could solve problems like maze navigation and robot locomotion more effectively than fitness-based evolution. Their work challenged the prevailing assumption that objective functions are necessary for progress in evolutionary computation. Novelty search has since been applied in various domains, including game playing, neural architecture search, and generative art.
The algorithm's simplicity and effectiveness made it a landmark contribution to the field. It also inspired subsequent research on diversity maintenance and quality-diversity algorithms, which combine novelty with performance metrics.
POET
POET (Paired Open-Ended Trailblazer) is an algorithm that Lehman co-created with Stanley and others, introduced in a 2020 paper. POET generates a series of increasingly complex environments and agents that learn to solve them. It uses a co-evolutionary approach where both the environments and the agents evolve, allowing the system to create its own curriculum of challenges.
POET is notable for its open-endedness: it does not have a fixed final goal but instead continuously generates new problems that are just beyond the current capabilities of the agents. This enables the discovery of behaviors that might not arise in a fixed environment. POET has been applied to tasks like maze navigation and bipedal locomotion, where it produced agents that could handle a wide range of terrains.
The algorithm represents a significant step toward creating AI systems that can learn and adapt indefinitely, without human intervention. It has been cited as an example of how open-ended learning could lead to more general and robust intelligence.
Research at OpenAI
Lehman joined OpenAI in 2020, where he has worked on various projects related to large-scale machine learning. His contributions include research on reinforcement learning, evolutionary strategies, and the training of large language models. He has been involved in efforts to improve the safety and robustness of AI systems, as well as to understand their emergent capabilities.
At OpenAI, Lehman has collaborated with other researchers on topics such as curriculum learning, reinforcement learning from human feedback, and the use of evolutionary algorithms for neural architecture search. His work has influenced the development of generative AI models, including those based on the transformer architecture.
Lehman has also written about the philosophical implications of open-endedness, arguing that it may be a key ingredient for achieving artificial general intelligence. He has given talks and published essays on the importance of exploration and diversity in AI research.
Key Papers and Publications
Lehman has authored or co-authored numerous influential papers. Some of the most notable include:
- "Abandoning Objectives: Evolution Through the Search for Novelty Alone" (2008), which introduced novelty search.
- "Evolving a Diversity of Virtual Creatures through Novelty Search and Local Competition" (2011), which extended novelty search to create diverse morphologies.
- "POET: End-to-End Learning of Open-Ended Environments" (2020), which introduced the POET algorithm.
- "Machine-generated knowledge and the future of AI" (2022), a paper exploring how AI can generate new knowledge.
- "The Surprising Creativity of Digital Evolution" (2018), a review of how evolutionary algorithms can produce creative solutions.
His publications have appeared in top venues such as the Proceedings of the National Academy of Sciences, the International Conference on Learning Representations (ICLR), and the Genetic and Evolutionary Computation Conference (GECCO).
Awards and Recognition
Lehman's work has been recognized with several awards. He received the Outstanding Paper Award at GECCO in 2008 for his novelty search paper. His POET paper was nominated for best paper at ICLR 2020. He has also been invited to speak at major conferences and workshops on evolutionary computation and AI.
In addition to academic recognition, Lehman's research has been widely covered in the media, including articles in Wired, MIT Technology Review, and Quanta Magazine. His ideas about open-endedness have sparked discussions about the future of AI and its potential to innovate beyond human expectations.
Impact and Legacy
Joel Lehman's contributions have had a lasting impact on the fields of evolutionary computation and artificial intelligence. Novelty search and POET are considered foundational works in the area of open-ended evolution, which seeks to create AI systems that can generate their own goals and continuously improve. His research has inspired a new generation of scientists to explore diversity-based methods and to question the necessity of objective functions.
Lehman's work has also influenced practical applications. For example, novelty search has been used in game design to create non-player characters with varied behaviors, and POET has been applied to robotics and automated curriculum generation. His ideas are increasingly relevant as AI systems become more complex and are expected to operate in dynamic, unpredictable environments.
As of 2025, Lehman continues to work at OpenAI, where he contributes to the development of advanced AI models and explores the frontiers of open-ended learning. His research remains at the forefront of efforts to build AI that is not only capable but also creative and adaptable.
Personal Life and Interests
Outside of research, Lehman is known for his interest in the philosophy of mind and the nature of creativity. He has written blog posts and essays on these topics, often drawing connections between biological evolution and artificial intelligence. He is also an advocate for open science and has made many of his papers and code publicly available.
Lehman is based in the San Francisco Bay Area, where OpenAI is headquartered. He enjoys hiking and playing board games, activities that he says inspire his thinking about exploration and strategy.
See Also
References
- Lehman, J., & Stanley, K. O. (2008). Abandoning Objectives: Evolution Through the Search for Novelty Alone. GECCO.
- Lehman, J., & Stanley, K. O. (2011). Evolving a Diversity of Virtual Creatures through Novelty Search and Local Competition. GECCO.
- Wang, R., Lehman, J., et al. (2020). POET: End-to-End Learning of Open-Ended Environments. ICLR.
- Lehman, J., et al. (2018). The Surprising Creativity of Digital Evolution. PNAS.
- Lehman, J., et al. (2022). Machine-generated knowledge and the future of AI. arXiv.