Learning, Incentives, and Optimization
for Networked Systems

Our research examines the role of incentives in allowing intelligent devices to cooperatively run applications requiring network and computing resources. Recently, we have focused on applications in networked machine learning and networked communities. We also maintain pages on specific projects, e.g., see here for an overivew of work done under Carlee's NSF CAREER award.

Smart Data Pricing

Initiatives like edge and fog computing are moving applications to an ever-larger array of devices in the Internet-of-Things, ranging from cloud servers to smartphones to low-power sensors. Yet relatively little work has examined whether this explosion is economically sustainable. Our work aims to do so by developing economic theories of the markets for data and computing resources that underlie the success of the Internet, and analyzing the effectiveness of existing pricing mechanisms. A common challenge is the dynamics of user demands over time, which often leads to dynamic pricing mechanisms in which users can pay for dynamic amounts of resources that match their demands at different times. Real-life examples include burstable instances for cloud computing and supplemental discounts for mobile data plans. Such dynamic pricing plans can be generalized to customized plans, in which users pay for resources according to their specific needs. Pricing mechanisms that aggregate services from multiple providers, such as Google's Fi data plan, can give users even more flexibility, though they are more challenging to analyze and offer as providers must have an incentive to offer such aggregated services.

Pricing data networks

Pricing compute resources

Non-monetary cooperation incentives

Networked Learning

Many emerging applications that utilize paradigms like edge and fog computing are based on machine learning and artificial intelligence, which have achieved remarkable success in a range of areas like vision, gaming, and autonomous vehicles. However, running conventional ML algorithms on networks of (possibly unreliable) devices may induce complex cost and performance tradeoffs that are not yet well understood. Our work develops and analyzes new algorithms for such networked settings, extending our work on dynamic pricing for network and computing resources to focus on the specific resource requirements and performance metrics of machine learning applications.

Distributed model training

Making ML inferences

Networked Communities

The pervasiveness of Internet-connected devices has enabled new types of physical and virtual communities in which users can proactively cooperate with each other to accomplish shared or individual goals. Our work examines how users in cities and online platforms can be incentivized to productively interact with each other. Some of our most recent work considers cooperation between vehicles in urban settings and user interaction in the virtual communities created by MOOCs (massive open online courses). Developing incentive mechanisms in these settings is particularly difficult since there are large numbers of users present that may interact in complex ways. For example, vehicles can move towards many different locations in a city, and each direction of movement will have a different impact on road congestion and other vehicles' chosen movements. Our ongoing work uses machine learning approaches to handle these complexities.

Incentivizing user mobility

Online courses