Georgia Institute of Technology College of Sciences

We consider a version of the knapsack problem in which an item size is random and revealed only when the decision maker attempts to insert it. After every successful insertion the decision maker can choose the next item dynamically based on the remaining capacity and available items, while an unsuccessful insertion terminates the process. We propose a new semi-infinite relaxation based on an affine value function approximation, and show that an existing pseudo-polynomial relaxation corresponds to a non-parametric value function approximation. We compare both theoretically to other relaxations from the literature and also perform a computational study. Our new relaxation provides tight bounds over a variety of different instances and surprisingly becomes tighter as the number of items increases. Joint work with Daniel Blado (ACO) and Weihong Hu (ISyE).

We resolve an open question from (Christiano, 2014b) posed in COLT'14 regarding the optimal dependency of the regret achievable for online local learning on the size of the label set. In this framework the algorithm is shown a pair of items at each step, chosen from a set of n items. The learner then predicts a label for each item, from a label set of size L and receives a real valued payoff. This is a natural framework which captures many interesting scenarios such as collaborative filtering, online gambling, and online max cut among others. (Christiano, 2014a) designed an efficient online learning algorithm for this problem achieving a regret of O((nL^3T)^(1/2)), where T is the number of rounds. Information theoretically, one can achieve a regret of O((n log LT)^(1/2)). One of the main open questions left in this framework concerns closing the above gap. In this work, we provide a complete answer to the question above via two main results. We show, via a tighter analysis, that the semi-definite programming based algorithm of (Christiano, 2014a), in fact achieves a regret of O((nLT)^(1/2)). Second, we show a matching computational lower bound. Namely, we show that a polynomial time algorithm for online local learning with lower regret would imply a polynomial time algorithm for the planted clique problem which is widely believed to be hard. We prove a similar hardness result under a related conjecture concerning planted dense subgraphs that we put forth. Unlike planted clique, the planted dense subgraph problem does not have any known quasi-polynomial time algorithms. Computational lower bounds for online learning are relatively rare, and we hope that the ideas developed in this work will lead to lower bounds for other online learning scenarios as well. Joint work with Pranjal Awasthi, Moses Charikar, and Andrej Risteski at Princeton.

Many statistical physics models are defined on an infinite lattice by taking appropriate limits of the model on finite lattice regions. A key consideration is which boundary to use when taking these limits, since the boundary can have significant influence on properties of the limit. Fixed boundary conditions assume that the boundary cells are given a fixed assignment, and free boundary conditions allow these cells to vary, taking the union of all possible fixed boundaries. It is known that these two boundary conditions can cause significant differences in physical properties, such as whether there is a phase transition, as well as computational properties, including whether local Markov chain algorithms used to sample and approximately count are efficient. We consider configurations with free or partially free boundary conditions and show that by randomly extending the boundary by a few layers, choosing among only a constant number of allowable extensions, we can generalize the arguments used in the fixed boundary setting to infer bounds on the mixing time for free boundaries. We demonstrate this principled approach using randomized extensions for 3-colorings of regions of Z2 and lozenge tilings of regions of the triangle lattice, building on arguments for the fixed boundary cases due to Luby et.al. Our approach yields an efficient algorithm for sampling free boundary 3-colorings of regions with one reflex corner, the first result to efficiently sample free boundary 3-colorings of any nonconvex region. We also consider self-reducibility of free boundary 3-colorings of rectangles, and show our algorithm can be used to approximately count the number of free-boundary 3-colorings of a rectangle.

How much of space can be filled with pairwise non-overlapping copies of a given solid? This is one of the oldest problems in mathematics, intriguing since the times of Aristotle, and remaining remarkably elusive until present times. For example, the three-dimensional sphere packing problem (posed by Kepler in 1611) was only solved in 1998 by Ferguson and Hales. In this talk, I will provide some historical and modern applications of geometric packing problems, and I will introduce a methodology to derive upper bounds on the maximal density of such packings. These upper bounds are obtained by an infinite dimensional linear program, which is not computationally tractable. However, this problem can be approximated by using tools from sums of squares relaxations and symmetry reduction (harmonic analysis and representation theory), leading to rigorous computational upper bounds on the density. Time permitting, I will present ongoing work with Maria Dostert, Fernando de Oliveira Filho and Frank Vallentin on the density of translative packings of superspheres (i.e., ell_p balls). This is an introductory talk: no previous knowledge of sums of squares relaxations or symmetry reduction is assumed.