Title: Sequential Decision Making at Scale: Theory and Applications
Date: June 5, 2026
Time: 1 PM ET
Location: ISyE Groseclose Building, Room 304
Meeting Link: https://gatech.zoom.us/my/neelkamalbhuyan

Neelkamal Bhuyan
Machine Learning PhD Candidate
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology

Advisor:
Dr. Debankur Mukherjee, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Committee:
•    Dr. Debankur Mukherjee, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
•    Dr. Siva Theja Maguluri, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
•    Dr. Vidya Muthukumar, School of Electrical and Computer Engineering, Georgia Institute of Technology
•    Dr. Adam Wierman, Department of Computing and Mathematical Sciences (CMS), California Institute of Technology
•    Dr. Souvik Dhara, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Abstract:
This thesis bridges the gap between traditional online optimization and adjacent decision-making literatures to solve complex resource allocation problems in large-scale systems. Novel methodologies are introduced by integrating stochastic optimal control, spectral perturbation analysis, decentralized optimization and game-theoretic frameworks into online decision-making problems that face massive reconfiguration penalties. Algorithms derived from these techniques yield robust, distribution-agnostic performance guarantees across centralized and multi-agent settings. Furthermore, these approaches are shown to achieve substantial real-world gains in critical domains like decentralized dynamic network slicing and learning-augmented scheduling for AI workloads.

Spread across six chapters, the first three focus on online convex optimization with unbounded metric switching costs, with the fourth chapter demonstrating operational gains in the network slicing domain. The next two chapters shift focus to the online workload scheduling problem, first presenting state-of-the-art robustness guarantees and then applying them to public cloud services to realize real cost savings. Together, these results present a principled approach for handling reconfiguration overheads in dynamic large-scale systems.