Sargis Vardanyan, Ashot Harutyunyan

Combining Classical Rule Learners with DSGD

Bio:

Sargis Vardanyan is a researcher at IIAP (NAS RA) and a junior ML engineer, currently completing an M.Sc. at the American University of Armenia. He holds a B.Sc. in Applied Mathematics & Computer Science from Lomonosov Moscow State University, where his undergraduate thesis explored reinforcement‑learning–based UAV control for target tracking. He used actor–critic RL systems for drone tracking. Also, he has built lightweight diffusion pipelines for image generation that run on consumer laptops and many other pet projects. His current work designs and implements an explainable Dempster–Shafer–based evidence‑aggregation framework that fuses automated rule induction (RIPPER/FOIL) with learnable mass parameters, reduced‑error post‑pruning, and simple rule deduplication for compactness.

 

 

 

Ashot Harutyunyan is currently leading the ML Lab at Yerevan State University (YSU). He was previously conducting data science research at VMware for intelligent cloud administration. He received his MS and Ph.D. degrees in applied mathematics at YSU and in information theory at the Institute for Informatics and Automation Problems of the National Academy of Sciences of Armenia in 1994 and 1997, respectively. He is an Alexander von Humboldt Fellow from the University of Duisburg-Essen, Germany, and a lecturer at YSU and the American University of Armenia in Statistics, ML, and information theory. His research interests and subjects of publications include applications of ML and design of algorithms, foundations of data transfer and compression, and related areas such as hypothesis testing. He is also an inventor with circa 80 US patents in AI operations for automated management of data center environments.

 

Description of the Talk:

We investigate whether DSGD on top of automatically induced rules improves interpretability without sacrificing accuracy. Rules come from RIPPER/FOIL and simple statistical heuristics; we compare raw rule sets vs DSGD-refined rules and against tree-based methods on tabular datasets.