Gaming the Research: Reinforcement Learning Changing Data Evaluation Challenges

BUFFALO, N.Y., August 3, 2021 — Advances in artificial intelligence, specifically reinforcement learning, are proving beneficial to accelerating the pace of data-intensive challenges, allowing scientists to get more work done and speeding up discovery.

During the American Crystallographic Association’s 71st annual meeting, Structural Science Awakens, which will be held virtually July 30-Aug. 5, Daniel Olds, from Brookhaven National Laboratory, will outline how reinforcement learning agents enhance work done with the Pair Distribution Function (PDF) beamline instrument at the National Synchrotron Light Source II (NSLS-II) facility. His presentation, “Beamline Science — The Video Game: How Reinforcement Learning and Artificial Intelligence are Changing Operations at Large-Scale User Facilities,” will be held Tuesday, Aug. 3 at 1:20 p.m. Eastern.

The methods used by Olds and his team with reinforcement learning (RL) are techniques often used in video games. By applying gamification to scientific processes, RL agents can learn as they are used in experiments, in effect, leveling up their rates of discovery as they work.

“As they interact with their environment, they can continue to grow and expand their collective experiences or memory,” Olds said. “These RL agents then reflect upon their collective memory to learn and adapt their behavior. This means that, just like a video game AI, they can slowly adapt their policies overtime to adjust to new or changing environments.”

Olds is using his trained RL agents at NSLS-II to accelerate the analysis of data-heavy measurements, hurrying up the pace of measurements and getting more user science done, a very significant improvement.

“But, of course, AI has a lot more to offer in regard to machine operation, automation, and data analysis,” he said. “In all of these cases, we are thinking about how the AI can augment and assist the scientists and engineers. In much the same ways computers allowed us to accelerate what we do as scientists, AI is going to be one of the tools we are all going to be using in the future in almost all aspects of science.”

The goal is not so much to make an AI that would fully automate the entirety of a beamline experiment, but rather, to act as a digital assistant to the researchers.

###

———————– MORE MEETING INFORMATION ———————–

IMPORTANT LINKS

Main meeting website: https://www.acameeting.com/ 
Technical program: https://www.acameeting.com/2021-program 
Event calendar: https://acas.memberclicks.net/index.php?option=com_jevents&Itemid=1&task=month.calendar

PRESS REGISTRATION FOR MEETING SESSIONS

We will grant free registration for credentialed and professional freelance journalists who wish to attend the meeting sessions. If you are a reporter and would like to attend, contact the AIP Media Line at [email protected]. We can also help with setting up interviews and obtaining images, sound clips or background information.

ABOUT AMERICAN CRYSTALLOGRAPHIC ASSOCIATION

The American Crystallographic Association, Inc. is a nonprofit, scientific organization of more than 1,000 members in more than 35 countries. The ACA was founded in 1949 through a merger of the American Society for X-Ray and Electron Diffraction (ASXRED) and the Crystallographic Society of America (CSA). The objective of the ACA is to promote interactions among scientists who study the structure of matter at atomic (or near atomic) resolution. These interactions will advance experimental and computational aspects of crystallography and diffraction. Understanding the nature of the forces that both control and result from the molecular and atomic arrangements in matter will help shed light on chemical interactions in nature.

###