At SODALab, we work broadly on algorithms and decision-making under uncertainty, particularly in systems involving online and multi-agent interactions. A central theme of our research is understanding how algorithmic performance scales with structural properties arising from three interacting dimensions of a decision system: the system itself (e.g., objectives, constraints, and strategic interactions), the environment generating uncertainty (e.g., sequential, adversarial, and stochastic variability), and the information interface between the two (e.g., predictions, offline data, and feedback acquired through interaction). On the applied side, our work is primarily driven by complex resource allocation and scheduling problems in large-scale systems such as online advertising, cloud computing, electrical grids, and LLM inference systems. For more details, explore examples of our recent projects below.
Online Algorithms
Research in online algorithms studies optimization and decision-making problems in which data are revealed incrementally over time and decisions must be made sequentially, often without full knowledge of the future. Unlike offline optimization, where all information is available upfront, online algorithms must operate under uncertainty while achieving strong competitive guarantees. Such problems arise naturally in modern systems such as online advertising, recommendation systems, financial platforms, cloud computing, and emerging AI infrastructures.
Our research in this area focuses broadly on the foundations of online algorithms and decision-making under uncertainty. In particular, we study how structural properties arising from objectives, constraints, and information shape the performance and limitations of online algorithms, with applications to online markets, cloud systems, and LLM inference systems.
Recent Highlights
Fairness in Online Decision Systems
Fairness is a difficult topic to talk about, as perspectives on fairness often vary across individuals and depend on the context and application scenarios. Nevertheless, addressing fairness in modern decision systems has far-reaching implications, particularly in high-stakes domains such as healthcare, education, and digital platforms. While classical notions such as envy-freeness, proportionality, and maximin share have been extensively studied in offline settings, fairness becomes significantly more challenging in online environments due to uncertainty, irrevocable decisions, and more importantly, the need to continuously coordinate interacting agents over time.
Our research in this area focuses broadly on fairness in online decision systems. In particular, we investigate how fairness objectives interact with structural properties such as group arrivals, predictions, repeated interactions, and multi-agent coordination. More broadly, we aim to develop principled frameworks for navigating trade-offs between fairness, robustness, adaptivity, and efficiency in sequential decision-making environments.
Recent Highlights
Learning-Augmented Algorithms
Learning-augmented algorithms, also known as algorithms with predictions, represent a recent paradigm that integrates machine-learned predictions into classical algorithmic frameworks to enhance practical performance while retaining provable worst-case guarantees. Unlike purely data-driven methods, these algorithms incorporate predictions to guide decision-making, yet are explicitly designed to remain robust even when predictions are highly inaccurate. This paradigm bridges the gap between worst-case analysis and empirical performance, combining the reliability of traditional algorithm design with the adaptability of machine learning.
Our research in this area investigates how predictions can be leveraged to surpass the worst-case guarantees of conventional algorithm design, particularly in online settings, with respect to metrics such as efficiency, fairness, and risk. We adopt both theoretical and applied perspectives, with applications ranging from electric vehicle charging and cloud computing to network caching and other decision-making problems under uncertainty.
Recent Highlights
Algorithms, Incentives, and Sustainable Systems
Modern decision systems are increasingly shaped by large-scale coordination problems involving uncertainty, incentives, and real-time decision-making. As demand for reliable and sustainable energy grows, particularly with the rise of AI data centers, there is an urgent need for algorithmic and economic mechanisms that align individual decisions with broader system-level objectives such as efficiency, reliability, and carbon sustainability. These challenges arise naturally in modern energy infrastructures involving distributed energy resources, electric mobility, and AI computing systems.
Our research in this area integrates tools from optimization, game theory, and machine learning to develop robust and adaptive algorithms for energy and sustainability applications. In particular, we study how incentives, uncertainty, and system constraints interact in resource allocation and coordination problems arising in power systems, electric vehicle charging, and data-center-scale infrastructures.
