with Aditya Nori, Akash Lal and Sriram Rajamani
We apply the technique of Counterexample Guided Abstraction Refinement (CEGAR) widely used in software model checking and theorem proving to accelerate relational inference. The resultant algorithm can be built upon existing relational inference tools, and significantly improves their performance.
We apply ideas from Kleinberg’s construction of small world networks to construct a set of random options that guarantee a logarithmic bound on the number of decisions required to reach a maximal value state for any particular task (i.e. reward function). Options generated perform comparably to existing state of the art option generation techniques. We also show that our options can be significantly cheaper to learn than those generated by existing techniques
We study and propose two novel techniques for discovering spatial and temporal abstraction in Markov Decision Process (MDP), homomorphic filters and small world options respectively. Our techniques are model-free, and perform well on standard examples.
with Kirtika Ruchandani and Balaraman Ravindran
We look at low-order approximations for the correlated topic model to collapse the non-conjugacy and retrieve a highly parallelisable inference procedure. Our solution turns out to be a rediscovery of prior work by Eric Xing. [article]
with Ramya Korlakai Vinayak and Gaurav Raina
We study two of the simplest quadratic population models with two distinct timescales, show that they capture the range of phenomenon from fixed points, to limit cycles, to chaos and analytically characterise the same using Hopf bifurcation theory. [report]