AI Program Solves Decades-Old Math Problems Using Reinforcement Learning

AI program plays the long game to solve decades-old math problems

In a groundbreaking development, a team of mathematicians led by Caltech’s Sergei Gukov has developed a new AI program that can solve math problems requiring extremely long sequences of steps. This program has the ability to think farther ahead than even advanced programs like AlphaZero, which famously triumphed over the best computerized chess engines in 2017.

The team’s AI program, developed using reinforcement learning, has made significant progress in solving families of problems related to the Andrews–Curtis conjecture, a challenging group theory problem proposed 60 years ago. While the main conjecture remains unsolved, the team has disproved families of potential counterexamples that have been open for decades, providing new insights and building intuition about the original problem.

The AI program, trained to come up with “super moves” or outliers, has shown remarkable success in navigating complex math problems that require thousands, millions, or even billions of steps. This innovative approach has already garnered attention in the math community, with new mathematicians joining the team to further explore the capabilities of the AI program.

Rather than relying on large amounts of computing power, the team’s strategy focuses on finding clever tricks and strategies that can be easily reproduced on small-scale computers. This approach has led to significant improvements in an area of math that had seen relatively slow progress, showcasing the potential of AI in tackling some of the toughest math problems.

The team’s work represents a significant step forward in the intersection of AI and mathematics, demonstrating the power of machine learning algorithms in solving complex and longstanding mathematical challenges. As they continue to push the boundaries of what AI can achieve in the realm of mathematics, the possibilities for future breakthroughs are endless.