AI-Assisted MemComputing: Generating Quality Metrics for Hyperparameter Optimization (Part 2)

MemComputing machines are dynamical systems that leverage memory (time non-locality) and collective behavior to solve combinatorial optimization problems. Their equations of motion depend on a few hyperparameters that can be tuned to further improve performance. In this talk, Part 2 of AI-Assisted MemComputing, we discuss how AI agents can not only optimize these parameters but also generate new quality metrics that better characterize solution dynamics. In particular, we query a Large Language Model to propose new metric functions that take various dynamical variables as inputs and whose output correlates well with the quality of a given set of parameters. We have developed a 2-agent pipeline in which a researcher agent, which proposes individual metrics and analyzes their effectiveness, and a supervisor agent, which provides suggestions based on large-scale trends in past experiments, interact with one another. We find that these agents produce metrics that are physically insightful and yield promising performance.