Eklavya: Research Projects

Fair Division of Indivisible Items
(August 2021  present)
Topics: cooperative game theory.
Supervisor:
Prof. Jugal Garg, ISE, UIUC.

(Jan 2020  April 2021)
Topics: approximation algorithms, bin packing, online algorithms.
Supervisor:
Prof. Arindam Khan, CSA, IISc Bangalore.
 Designed approximation algorithms for rectangle bin packing when items
are skewed, i.e., each item either has small width or small height.
 Designed approximation algorithms for a variant of bin packing that
generalizes geometric bin packing and vector bin packing.
 Designed an approximation algorithm for ddimensional geometric bin packing
when items are allowed to be rotated. This algorithm gives the bestknown
approximation factor for d ≥ 3.
 Worked on the online knapsack problem in the randomorder model.
Obtained hardness results and improved algorithms for some special cases
(profit=size and profit=1).

(Sept 2017  Dec 2017)
Topics: computer networks, network security, SDN.
Supervisor:
Prof. Vishal Gupta, BITS Pilani.
 Studied DNSrelated DoS attacks and SoftwareDefined Networking (SDN).
 Devised a new mechanism for mitigating DNS amplification attacks,
which uses a set of geographicallydistributed SDN routers.
 I presented a paper on it at
ICACCI in September 2018.

(Jan 2018  June 2018)
Topics: neural networks, machine learning, big data.
 Trained a neural network from almostraw data to estimate
the probability of a creditcard applicant defaulting.
 The data was in a unique format, so a custom neural network architecture was devised.
 The neural network's performance was at par with the model then in production,
which was tuned over many years and utilized several complex handengineered features.

(Nov 2017  Jan 2018)
Topics: machine learning, algorithms, math.
Supervisor:
Prof. Surekha Bhanot, BITS Pilani.
 Invented a clustering algorithm, which I named CTmeans.
It is an approximation to Cmeans fuzzy clustering.
It uses KDtrees to reduce running time.
 Mathematically proved its convergence and approximation guarantees.
 Implemented
the algorithm and benchmarked its performance on different datasets.
It was not significantly faster in practice and its applicability was limited.