Projects
My group works at the interface of game theory, economics, control theory, optimization, (machine/statistical) learning, and
artificial intelligence.
Motivated by learning-enabled
intelligent systems (e.g., intelligent infrastructure systems, platform based markets
supporting e-commerce/healthcare/transportation/logistics, etc.), our
research seeks algorithms and methodological innovations that have broad applicability.
Learning-based markets and other forms of automated/algorithmic decision
making are now commonplace in a variety of critical areas of our economy.
And, algorithms are increasingly being incorporated into complex systems where they must learn in the presence
of competition and execute sequences of decisions in the face of scarcity of
resources, under uncertainties that influence the ability to robustly draw future
inferences and satisfy downstream objectives, and in the presence of
strategically generated (even adversarial) data. The research in my group seeks to
develop algorithms for everything from inference to influence in non-stationary,
uncertain, and competitive environments. Below are short descriptions of some
the projects that align with this theme.
A common thread across core research projects is the development of
performance guarantees for coupled
decision-making agents and/or learning
algorithms and seeking such guarantees even in non-stationary environments.
Applications Areas: human-machine
interaction & autonomy, online platforms/data markets, societal-scale
systems (e.g., intelligent transportation systems).
Funding
This work is largely funded by the National Science Foundation ((CRII) CNS-1656873, (SCC) CNS-1736582, (CPS) CNS-1836819, (CPS) CNS-1931718,
(RI)
IIS/CSE-1907907,
(CAREER) CNS-1844729, SCC Track
1 CNS-1952011), and the
Office of Naval Research (ONR YIP)
Past funding: (CPS) CNS-1634136,
(US-Ignite) CNS-1646912
Awards & Recognition
|
Certifiable Reinforcement Learning |
||||||||||
| Many of the same kinds of
uncertainties (particularly, those arising from human
decision-makers) arise in a number of applications where research
into the efficacy of reinforcement learning approaches to control
and automation are being explored. For example, in (semi-)autonomous driving and robotic surgery applications humans are directly coupled with another autonomous agent and real-time decisions are being made often over short horizons where there is little time for cogitation and risk or reference points (e.g., past experiences) are the basis for decisions. We are developing algorithms for
reinforcement learning with certificates of performance with assessment on
benchmark problems. Applications include intelligent infrastructure
(societal-scale CPS) and human-in-the-loop systems. [joint with Sam
Burden (UW); Sam Coogan (GaTech)] |
||||||||||
|
||||||||||
Recent Papers
:
|
||||||||||
|
Adaptive Mechanism Design and Multi-Sided Markets |
||||||||||
| In situations of asymmetric information, we are designing incentive or information shaping schemes for changing user behavior via algorithmic mechanisms, with provable guarantees on performance, that leverage behavioral models capturing salient features of the human decision-making process. We take motivation from a number of applications including peer review (bidding and reviewer-paper matching), online labor markets (proposals and reverse selection), and incentive/information design in mobility applications including on-street parking, micro-vehicle sharing, concierge service companies (e.g., delivery on-demand), etc. [collaborators: K. Jamieson, UW; L. Jain, UW; Nihar Shah, CMU] | ||||||||||
|
||||||||||
Recent Papers:
|
||||||||||
|
Game Theory for Building Foundational ML Tools |
||||||||||
| We often seek machine learning algorithms with guarantees on robustness to adversarially generated examples or strategically generated data. Increasingly the formulation of such problems takes the form of a game. For example, everything from adversarial learning to fair classification has been formulated as a minmax optimization problem or zero-sum game. Yet, many of the known results for such problems often assume a lot of problem structure (e.g., convexity of costs or full information) which is not present in machine learning problems, where non-convexity and limited information are quite common. In this line of research we seek to understand the optimization landscape for such problems and design algorithms that not only lead to game theoretically meaningful equilibria but learning algorithms with good performance and robustness guarantees. | ||||||||||
 |
||||||||||
Recent Papers:
|
||||||||||
|
Algorithmic Competition & Cooperation |
||||||||||
| A number of machine learning problems are begin cast as games in order to learn more robust, better performing algorithms. On the other hand, the rise of multi-agent learning (e.g., AlphaGo) in non-stationary environments as well as competition between algorithms (e.g., in financial markets or online marketplaces) exposes the need for better analysis tools for the interaction between learning agents/algorithms. With these two observations in hand, our goal is to derive performance guarantees for classes of (competitive and cooperative) multi-agent learning algorithms in non-stationary environments. Such tools are useful for both analysis of the behavior of coupled learning agents as well as synthesizing mechanisms for shaping agent behavior. In addition, we seek to draw connections between dynamical systems theory (in particular, types of limiting behavior beyond equilibria) and learning dynamics. | ||||||||||
 |
||||||||||
Recent Papers:
|
||||||||||
|
Applications |
||||||||||
| Intelligent Transportation: Curbside & Real-Time Fleet Management | ||||||||||
| A major source of motivation in some of the research we do are problems in urban mobility including curbside management and real-time fleet management for public transit. The emergence of novel digitally mediated markets for movement of goods and people as well as on-demand door-to-door services has lead to a number of interesting challenges requiring the careful design of market mechanisms that do not introduce new unintended outcomes such as unfair treatment of socioeconomic groups or increased congestion. Furthermore, a number of classical problems in transportation such as curbside management are revealing themselves to be very interesting problems to revisit with a modern perspective due to the availability of data and the new transit mechanisms such as ride-sharing, concierge services, and micro-vehicles (e.g., scooters, bikes, etc.). | ||||||||||
|
||||||||||
Curbside Management in Seattle, WA.
Research in
this direction has been conducted in collaboration with the Seattle
Department of Transportation (on curbside management) where we have had
impact on the way in which data is curated, analyzed and used to make
policy decisions from pricing to resource allocation to assessing curb
availability in development projects. We are developing methods of
modeling and predicting occupancy from paid occupancy (block level
paid transaction data indicating arrival and estimated departure of
paying customers) and sparse
groundtruth data from manual studies (walking around and counting cars)
and timelapse cameras.
This is challenging because the paid transactions
do not provide the full picture for occupancy and ground truth data is
sparse and expensive to collect. Typically one day per year per block is
collected in Seattle. To address this, we are developing
methods of actively (active learning via contextual pure exploration) selecting locations to target longer duration
studies at locations that are the most informative in terms of improving
the occupancy model quality.
|
| 
| 
| 
|
| Real-time Public Transit Fleet Management in Chattanooga and
Nashville, TN. Research on real-time
fleet management is in collaboration with CARTA (Chattanooga Area
Regional Transportation Authority), WeGo in Nashville TN, and Vanderbilt
University (Abhishek Dubey, project lead). In this line of work we seek
to design routing strategies for allocating micro-fleet
vehicles that meet real-time demand for public transit by feeding into
the fixed bus routes. More detail on this project can be found at
Smarttransit.ai.
|
|
| 
| ||
Recent Papers:
|
||||||||||
| | ||||||||||
| Human-Machine Interaction & Behavior Models of Humans | ||||||||||
|
||||||||||
Recent Papers:
|
||||||||||
| | ||||||||||
| Online Platforms | ||||||||||
|
||||||||||
Recent Papers:
|