The paper demonstrates how generative AI can improve upon existing techniques for solving optimization problems common in industrial settings.
Zapata Computing Holdings Inc. (“Zapata AI”) (Nasdaq: ZPTA), the Industrial Generative AI company, today announced that its foundational research on generator-enhanced optimization (GEO) has been published in the esteemed Nature Communications journal. The research, titled “Enhancing Combinatorial Optimization with Classical and Quantum Generative Models,” introduces Generator-Enhanced Optimization (GEO), a novel optimization method that leverages the power of generative modeling to suggest high-quality candidate solutions to complex optimization problems. It is the second Zapata AI paper on generative AI to be published in Nature Communications since December 2023.
The paper discusses our findings when we have tested GEO for financial portfolio optimization, finding that GEO performs competitively and often outperforms existing state-of-the-art optimization algorithms, which have been fine-tuned for decades. Portfolio optimization is a common problem among investors who aim to allocate their capital to maximize their returns for a given level of risk (or minimize their risk for a desired level of returns). Despite years of study, this problem remains a computational challenge for financial institutions that only becomes more challenging the more assets are involved. The GEO paper reflects the results of a pioneering effort to apply generative AI to portfolio optimization and other optimization problems.
“When a lot of business leaders think of Generative AI, they think of LLMs, but this research demonstrates one of the many ways generative AI can be applied to industrial problems beyond language tasks” said Christopher Savoie, CEO and co-founder of Zapata AI. “We believe generative AI is the next frontier in business analytics, whether that’s generating data for variables that couldn’t otherwise be measured or recommending better ways to solve optimization problems, as in this paper. It’s very exciting to see this continued validation of our work in generative AI and we’re immensely proud of the researchers involved.”
GEO has been applied to real-world industrial problems since the research paper was initially submitted to ArXiv in 2021. In 2022, GEO was used in work with BMW and the Center for Quantum Engineering at MIT to find more efficient manufacturing plant operating schedules, minimizing idle time between steps in the manufacturing process while meeting production targets. That research found that GEO tied or outperformed state-of-the-art optimization algorithms in 71% of problem configurations. More information on GEO can be found here.
Since GEO was first developed, Zapata AI has established a growing portfolio of quantum techniques for generative AI. For instance, Zapata AI researchers recently leveraged quantum-enhanced generative AI to generate viable cancer drug candidates for the first time. Quantum science could offer several advantages for enterprise problems, including compressing large, computationally expensive models; speeding up time-consuming and costly calculations; and generating more diverse, higher quality outputs for generative AI. More details on how quantum science can enhance generative AI can be found in a recent Zapata AI blog post.
“Our Nature Communications article reflects an early demonstration of how generative AI techniques inspired by quantum physics can be applied to solve optimization problems” said Mohammad Ghazi Vakili, a former post doc at Zapata AI who authored the paper along with Javier Alcazar, Can B. Kalayci, and Alejandro Perdomo-Ortiz. “It was impressive to see GEO go toe-to-toe or outperform algorithms that have been fine-tuned for decades. We expect to see more impressive results as quantum generative AI matures.”
