Joel Lehman: Open-Endedness and Evolution through Large Models

Origins

Joel Lehman's journey into AI began with a serendipitous encounter with a book on AI and game development, which introduced him to evolutionary algorithms and neural networks 1. His fascination with these concepts led him to explore the potential of neuroevolution, where evolutionary algorithms are used to evolve neural networks 1. Joel's interest was piqued by the idea that a simple algorithm could lead to complex and diverse outcomes, much like biological evolution 2.

An evolutionary algorithm evolving a brain. That's like how we got here maybe that's something that I could study.

--- Joel Lehman

This curiosity drove him to pursue graduate studies focused on these algorithms, shifting his focus from game programming to AI research 1.

   

Novelty

In the realm of AI, novelty search offers a unique approach to overcoming deception, a challenge where the path to a solution doesn't resemble the solution itself 3. Joel Lehman explains that novelty search circumvents this by encouraging exploration without a fixed objective, allowing for unexpected solutions to emerge 4. This method contrasts with traditional reinforcement learning, which can lead to dead ends when narrowly optimizing for specific goals 3.

Novelty is driven by diverging from where you've been and you know where you've been, and so you can know if something is different from where you've been.

--- Joel Lehman

By focusing on novel behaviors, AI systems can discover new paths that might otherwise be overlooked, demonstrating the potential of this approach in solving complex problems 5.

   

Innovation

Joel Lehman's work on evolutionary algorithms emphasizes the importance of open-ended innovation, where the search for novelty can lead to unexpected breakthroughs 6. By abandoning fixed objectives, these algorithms mimic the diverse and unpredictable nature of biological evolution, fostering a creative exploration of solutions 6. This approach aligns with the broader scientific process, where discoveries often serve as stepping stones for future innovations, even if their immediate applications are unclear 7.

It's sometimes unclear where the things you discover are going to lead.

--- Joel Lehman

Lehman's research highlights the potential of prioritizing interestingness over specific goals, allowing for a more dynamic and adaptable problem-solving process 7.