By Dave DeFusco
In the world of technology, making systems faster, smarter and more efficient often happens behind the scenes. David Sweet, an industry professor in the Katz School’s Graduate Department of Computer Science and Engineering, has developed a method for improving how computers optimize complex tasks in his paper, “Fast, Precise Thompson Sampling for Bayesian Optimization,” which was presented at NeurIPS, the largest AI conference, in December.
Dr. David Sweet's work has a surprising connection to many everyday technologies we use. Think about how streaming services recommend the perfect show. These systems test many possible approaches to recommendations and pick the one that works best for you. Bayesian optimization is a tool that helps computers do this efficiently by reducing the number of “tests” needed to find the best solution.
“The fewer experiments required, the faster and more cost-effective the process becomes,” he said. “This is critical for fields like healthcare, where optimizing a treatment plan quickly can save lives, or quantitative trading, where testing new ideas can lose money.”
In technical terms, Thompson Sampling (TS) is a method that selects the best options, or “arms,” by guessing which ones are most likely to succeed. Imagine an app trying to decide which ad you’d be most interested in—it uses TS to make educated guesses. While TS is great for simpler systems, it struggles in more complex scenarios. That’s where Dr. Sweet’s work comes in. He and his team created an advanced version called the Stagger Thompson Sampler (STS). This method is faster and more precise, meaning it can handle bigger challenges more effectively.
Here are some real-world applications:
Better Recommendations: From shopping apps to streaming services, these systems can now learn your preferences more quickly and suggest more relevant options.
Faster Drug Development: Scientists can test fewer configurations in their simulations, speeding up the creation of life-saving medications.
Greener Technology: Electric car charging schedules can by fine-tuned, speeding up charger stops and reducing wear on the battery.
Smarter AI: AI models can be made to learn faster, perform better and even consume less power.
Dr. Sweet's method isn’t just faster; it works in high-dimensional problems, which means it can tackle very complicated systems without breaking a sweat. STS has outperformed older methods or problems involving hundreds of variables, a scale that is both important and challenging in scientific and engineering problems.
“Even though you may never directly interact with Bayesian optimization or Thompson Sampling, these advancements are shaping the tools, apps and technologies we use every day,” said Dr. Sweet.
