Rutgers University: Department of MSIS Seminar Schedule
All the seminars are held on Fridays at 11am unless otherwise noted. For more information, please contact Mert Gurbuzbalaban.
Seminar Organizer(s): Mert Gurbuzbalaban and Thomas Lidbetter.
- Feb 24: No seminar due to Rutgers Research Day.
- March 3: Nikhil Vellodi (NYU)
- Time: 11am, Location: Newark 1WP-502
- Video streaming available at: New Brunswick, 100RR-4095
Title: Cheap Talk in Multi-Product BargainingAbstract: I study a game in which a buyer and a seller bargain over multiple products. The buyer has private valuations over the seller’s product range, and can communicate these through cheap-talk messages. In the benchmark model a fully revealing, welfare-maximizing equilibrium exists. However, a well-known refinement selects a non-empty set of partially informative equilibria, which I characterize fully. In particular, the set contains the ex-ante buyer optimal equilibrium. These results hold for a broad class of preferences, and rest on a natural trade-off faced by the buyer – reveal his type in order to secure his preferred good, or hide his type in order to secure a better price. As goods become less substitutable, the buyer reveals more information, whilst his share of the surplus changes non-monotonically.
- March 10: Mengdi Wang (Princeton)
- Time: 11am, Location: New Brunswick 100RR-3095
- Video streaming available at: Newark – 1WP-302
- Title: Stochastic First-Order Methods in Data Analysis and Reinforcement Learning
- Abstract: Stochastic first-order methods provide a basic algorithmic tool for online learning and data analysis In this talk, we survey several innovative applications including risk-averse optimization, online principal component analysis, and reinforcement learning. We will show that convergence analysis of the stochastic optimization algorithms provide sample complexity guarantees for the corresponding online learning applications.
- March 20 (Monday): Stan Uryasev (Florida)
- Time: 11am , Location: TBA
- Video streaming available at: TBA
- Title: Buffered Probability of Exceedance (bPOE): Mathematical Properties and Applications
- Abstract: (Joint work with Matt Norton and Alexander Mafusalov) The Probability of Exceedance (POE) is frequently used to measure uncertainties in outcomes. For instance, POE is used to estimate probability that assets of a company fall below liabilities. POE measures only the frequency of outcomes and ignores magnitude of outcomes. POE counts outcomes exceeds the threshold, and it “does not worry” about the amount by which each outcome exceeds the threshold. POE is lumping together all threshold exceedance events, potentially “hiding” quite large and very troublesome outcomes. Moreover, POE has poor mathematical properties when used to characterize discrete distributions of random values (e.g., when distributions are defined by observed historical data). POE for discrete distributions is a discontinuous function of control variables, making it difficult to analyze and optimize.This presentation discusses a new probabilistic characteristic called Buffered Probability of Exceedance (bPOE). With bPOE, it is possible to count outcomes close to a threshold value, rather than only outcomes exceeding the threshold. To be more precise, bPOE counts tail outcomes averaging to some specific threshold value. For instance, 4% of land-falling hurricanes in US have cumulative damage exceeding $50 billion (i.e., POE = 0.04 for threshold=$50 billion). It is estimated, that the average damage from the worst 10% of hurricanes is $50 billion. In terms of bPOE, we say bPOE=0.1 for threshold=$50 billion. bPOE shows that largest damages having magnitude around $50 billion have frequency 10%. bPOE can be considered as an important supplement to POE. We think that bPOE should be routinely calculated together with POE. This example shows that bPOE exceeds POE, which is why it is called Buffered Probability of Exceedance. The positive difference, bPOE-POE, can be interpreted as some “buffer.” The bPOE concept was recently developed as an extension of Buffered Probability of Failure (introduced by Rockafellar and Royset). bPOE has been derived from Conditional Value-at-Risk (CVaR) characteristic of uncertainty. Actually, bPOE is an inverse function of CVaR and it inherits a majority of the exceptional mathematical properties of CVaR (which is a so called “coherent measure of risk”). Similar to CVaR, minimization of bPOE can be reduced to convex and Linear Programming.We will discuss two applications of bPOE concept. The first application considers the Cash Matching of a Bond Portfolio. We minimize bPOE that assets exceed liabilities. The second application uses bPOE in data mining. Currently, the Area Under the Receiver Operating Characteristics Curve (AUC) is standardly used to evaluate classification models. AUC can be presented as the probability that some discrete random value is below zero. We explored so called Buffered AUC (bAUC) as a counterpart of the standard AUC.
- March 24: Steve Alpern (Warwick Business School)
- Time: 11am, Location: 1WP-1027 / 100RR-4095
- Title: The Rendezvous Search Problem
- Abstract: The rendezvous search problem asks how two (or more) unit-speed agents can minimize the expected time required to meet up when placed randomly in some search region, possibly a network. It was first posed informally in 1976 but did not receive attention in the literature until the 1990’s. It has since become an important problem in robotics, coordination theory and more recently in computer science. Many apparently simple problems are still unsolved.
- March 31: Eugene Feinberg (Stonybrook)
- Time: 11am, Location: TBA
- Title: Recent Developments in Markov Decision Processes Motivated by Inventory Control
- Abstract: The talk describes the recent progress in the theory of Markov Decision Processes (MDPs). This progress is motivated by inventory control applications of MDPs. The progress became possible because of recent generalizations of two facts in real analysis: Fatou’s lemma and Berge’s maximum theorem. In addition to inventory control applications, the talk describes new results on game theory and robust optimization.
- March 31: Serhat Aybat (Penn State)
- Time: 2pm, Location: 100R 5038, Video-conference: 1WP 921
- Title: A distributed ADMM-like method for resource sharing under conic constraints over time-varying networks
- Abstract: In this talk, a decentralized method for solving cooperative multi-agent optimization problems over a network of agents will be discussed, where only those agents connected by an edge can directly communicate with each other and the topology of the communication network may be time varying. The objective is to minimize the sum of agent-specific composite convex functions, i.e., each term in the sum is a private cost function belonging to an agent, subject to a conic constraint coupling all agents’ local decisions. An ADMM-like primal-dual algorithm is proposed to solve a saddle point formulation of the problem in a distributed fashion, where the consensus among agents is imposed on agents’ evaluations of the shadow price associated with the coupling constraint. We provide convergence rates in a) sub-optimality and infeasibility of agents’ local decisions, and b) consensus violation among agents’ shadow price estimates; examine the effect of underlying network topology on the convergence rate of the proposed decentralized algorithm; and show how to extend this method to handle directed communication networks. This optimization model abstracts a number of applications in machine learning, distributed control, and estimation using sensor networks. Joint work with Ph.D. student Erfan Yazdandoost Hamedani.
- Bio: Necdet Serhat Aybat is an assistant professor in the Department of Industrial and Manufacturing Engineering at Pennsylvania State University. He received his Ph.D. degree in operations research from Columbia University in 2011. His research focuses on developing first-order algorithms for large-scale convex optimization problems from diverse application areas, such as distributed optimization, robust matrix decomposition, convex regression, and compressed sensing. His current research is supported by two NSF grants: one on designing optimization algorithms that simultaneously learn missing model parameters of a given parametric model, and the other one on designing smart power grids through distributed optimization.
- April 14th: Melike Gursoy (Rutgers)
- Time: 11am, Location: 100RR 3095, Videoconference at: 1WP 302
- Title: Multi Server Queues and Task Completion Times in Random Environment
- Abstract: We first present some queueing models with modulated arrival and service processes. We show that these models demonstrate the stochastic decomposition property, and derive the stationary distribution of the number in the system. Then, we investigate the task completion time for customers in such a random environment with generally distributed service and down times and exponentially distributed up times. We consider both the preemptive repeat different (Replace) and preemptive repeat identical (Repeat) service policy and obtain Laplace transform of the completion time distribution. We discuss the conditions under which the task completion time has heavy tail.
- April 21st: Waheed Bajwa (Rutgers)
- Time: 11am, Location: TBA
- Abstract: TBA
- May 5th: Lisa Hellerstein (NYU)
- Time: 11am, Location: 1WP-1027 / 100RR-4095
- Abstract: TBA