Strategic learning of strategic behavior

Intelligent agents are designed to act in order to maximize their long-term reward. But agents rarely act in isolation -- they repeatedly interact with one another and influence each other's payoffs. And their behavioral history is often public. It is of interest to understand how to learn and make inferences from the behavioral history of others, as well as understand how to shape our own behavioral history to our advantage. Central to this goal is understanding how to play non-zero sum games beyond a worst-case sense. My current research aims to tackle these problems through the lens of active learning, online learning and game theory.

Here are brief descriptions of some of my completed and past projects.

The advantages and limitations of partial commitment

Joint work with Anant Sahai

Stackelberg play, in which some leader agent commits to a strategy or move first, is often socially advantageous over simultaneous play and arises in security and law enforcement games. But this solution concept assumes perfect demonstration, observation and belief in leader commitment; generally not the case in practice. We capture imperfect demonstration through a source-coding-based concept of partial commitment, and show that this concept can be thought of as a continuum between Stackelberg and simultaneous equilibrium. This work was presented at IEEE International Symposium on Information Theory, Aachen, Germany, 2017.

Current research includes capturing uncertainty in commitment observability and belief through a learning-theoretic model of reputation building.

Ex-post enforcement for dynamic spectrum sharing

Joint work with Anant Sahai


The current paradigm of ex-ante enforcement is easy to implement, but defines conservative exlusion zones for incumbent users of spectrum. This leads to inefficient use of spectrum and circumvents the crucial problem of spectrum sharing on a temporal basis. The case for an ex-post enforcement model, in which unlicensed users of spectrum are deterred from causing harm through a "spectrum jail" system has been argued in previous work. We delineate rights that can be guaranteed to a primary user of spectrum in terms of retained performance and show that compatible secondary users can use spectrum productively.

This work was presented at IEEE Symposium on Dynamic Spectrum Access Networks, Baltimore, 2017.

Past Projects

TV Whitespace in Canada and Australia

Joint work with Kate Harrison and Anant Sahai


My previous work focused on data-driven, quantitative analyses of existing unlicensed spectrum. The TV whitespaces, the incarnation-du-jour of dynamic spectrum access, are gaining attention worldwide. Until now, there has been no easy way to carry out data-driven calculations to estimate the whitespace opportunity under different spectrum allocation scenarios -- previously existing tools were closed source and not flexible enough to allow this.

Whitespace Evaluation SofTware(WEST), a modular and powerful open-source software developed and designed by Kate Harrison, former graduate student in EECS at UC Berkeley, supports whitespace availability calculations under highly versatile conditions. Use of the TV whitespaces was legalized in Canada at the beginning of 2015. Weeks after this development, we were able to use WEST to obtain novel results for whitespace availability in Canada and Australia under the FCC's and Industry Canada's regulatory sets for whitespace devices. We were also able to compare the older FCC ruleset with the newer Industry Canada ruleset in the USA, Canada, and Australia and study their differences in detail.

This work was presented at IEEE Symposium on Dynamic Spectrum Access Networks, Stockholm, 2015.

Effect of incentive auction on TV Whitespaces

Joint work with Angel Daruna, Vijay Kamble, Kate Harrison, and Anant Sahai


Privatization of spectrum allocation through auctions has been long advocated for to increase the efficiency of spectrum as an economic resource. The most recent manifestation of this has been the FCC's ongoing incentive auction that aims to clear up to 144 MHz of TV spectrum for LTE usage.

We analyzed optimally efficient re-allocations of TV stations for various spectrum clearing targets using satisfiability solvers, and the effect of such desirable auction outcomes on TV as well as TV whitespaces. This work was presented at IEEE International Conference on Communications, 2015.

I'm also a member of the UC Berkeley team for the upcoming DARPA Spectrum Collaboration Challenge, which aims to facilitate coordination and collaboration between separately designed systems to access spectrum productively.

Stochastic Decoding of LDPC Codes: Undergraduate Senior Thesis Project

Joint work with Andrew Thangaraj

Low-density-parity-check codes are important in applications requiring reliable and highly efficient information transfer. A communication system using LDPC codes consists of an encoder and decoder, and decoding is performed using the message-passing algorithm. We studied the performance of a stochastic decoder for LDPC codes whose computational complexity is much lower. We also developed density evolution equations for the stochastic decoder. We implemented modified versions of the stochastic decoder to optimally performing short LDPC codes over GF(q) and interpreted the results.

Constructing Low Coherence Matrices Using SL2(q)

Joint work with Matthew Thill and Babak Hassibi

Matrices with low coherence are important in compressed sensing, sphere decoding and quantum measurements among other applications. Good constructions have been developed for random matrices. We constructed deterministic low coherence matrices using concepts from group theory and representation theory. In particular, we looked at Abelian groups and the special linear group SL2(q).

This work was presented at IEEE International Conference on Acoustics, Speech, and Signal Processing, Florence, 2014.