At SODALab, we study algorithms and decision-making under uncertainty, with a particular interest in decision systems involving interacting people, institutions, and computational agents. A central theme of our research is to understand how algorithmic performance depends on three aspects of a decision system: the system itself, including its objectives, constraints, and strategic interactions; the environment, including adversarial and stochastic uncertainty, temporal patterns, and distributional structure; and the information available through predictions, offline data, and feedback acquired through interaction.
Much of our work can be viewed through the shared lens of resource allocation, motivated by applications such as online advertising, marketplaces, cloud computing, AI, and energy infrastructure. Across these settings, we ask which forms of structure, information, and incentives fundamentally change what algorithms can achieve. Explore some of our recent projects below.
Online Algorithms
Research in online algorithms studies decision-making problems in which information arrives over time and decisions must often be made irrevocably before the future is fully known. Such problems arise throughout modern decision systems, including online platforms, cloud computing, electrical grids, and large-scale computing infrastructure.
Our research asks how the capabilities and limitations of online algorithms change when classical assumptions are relaxed or enriched by additional structure. We study forms of decision flexibility such as bounded adaptivity, reusability, and cancellation; objectives involving risk and structured constraints; and additional information in the form of predictions, historical data, or feedback acquired through repeated operation. A central goal is to understand which features fundamentally change what can be achieved online, and to design algorithms that exploit them while retaining rigorous guarantees.
Recent Highlights
Sequential Multi-Objective Allocation
Shared-resource systems serve persistent users over time, where performance is naturally measured by a vector of utilities reflecting each participant’s cumulative benefit. In an inference service, for example, multiple applications draw on a shared compute pool—meaning that allocation decisions shape both total efficiency and individual fairness. Maximizing a single aggregate utility often leaves key participants poorly served. The challenge lies in balancing these vector components when future requests are unknown and early commitments restrict later opportunities.
Our research addresses multi-objective optimization in these settings through structured welfare functions that combine utility vectors to balance efficiency and fairness. We investigate how the choice of objective shapes algorithmic limits, and how these trade-offs depend on arrival patterns, available information, and decision flexibility.
Recent Highlights
Offline Data and Sequential Interaction
Modern decision systems often begin with historical data but must ultimately operate through sequential interaction with an uncertain environment. Offline data can improve initial decisions and reduce costly exploration, but it may be incomplete, biased, or collected under different policies. This raises a fundamental question: how should prior data and new information acquired through interaction be combined?
Our research studies the foundations of this offline-to-online transition, drawing on online learning, reinforcement learning, and statistical decision theory. We ask when historical data can be trusted, when new exploration is necessary, and how the quality and coverage of prior data shape what can be learned safely and efficiently during deployment.
Recent Highlights
Algorithms and Incentives in Infrastructure Systems
Modern infrastructure systems increasingly rely on algorithmic decisions to coordinate many interacting users, devices, and organizations under uncertainty. These systems must balance local incentives with broader objectives such as efficiency, reliability, and sustainability, often while operating in real time and under physical constraints. Such challenges arise across power systems, electric mobility, data centers, and other large-scale computing and energy infrastructures.
Our research studies how algorithms, incentives, and system constraints interact in these settings. We draw on optimization, game theory, and machine learning to design robust and adaptive mechanisms for resource allocation, coordination, and control, and to understand how individual behavior and institutional rules shape system-level outcomes.
