Seminar schedule: Year 2014

• Tuesday, march 18th 2014, Time:: 12:30

Vehicle routing and fleet sizing problems arising in servicing offshore oil and gas installations
Irina Gribkovskaia, Molde University College, Norway

• Monday January 20th 2014, Time: 12:30

Optimizing Personalized Predictions using Joint Models
Dimitris Rizopoulos, Department of Biostatistics, Erasmus University Medical Center, The Netherlands.

---

Top

Functional Data Analysis and Dimensionality Reduction

SPEAKER: Philippe Vieu
LANGUAGE: Anglès
PLACE: Edifici C5, Aula C5016, Campus Nord, UPC (veure mapa)
DATE: Divendres, 19 de desembre de 2014. Hora: 12:30
ABSTRACT: Functional data are, by nature, infinite dimensional data, and their analysis need necessarily specific attention to the dimensionality effects. Semi-parametric modelling and variable/model selection are two fields of modern Statistics which are appealing candidates for that purpose.

In a first attempt it will be shown along this talk, through the simple Single Functional Index Model, how semi-parametric modelling behaves for FDA. From a methodological point of view one will see how this model is rather flexible without being affected by the infinite-dimensionality effect.

In a second attempt, one will develop specific variable selection procedures for FDA. The methodology will take fully into account the continuous structure of the data, leading to rather low computational costs (compared with standard multivariate selection procedures).

The talk will be mainly methodological and centered around the presentation of these two functional methodologies. Then, it will end by the presentation of some benchmark real curves dataset analysis. From an applied point of view, it will be highlighted how the functional semi-parametric statistical procedures are combining good predictive power, low computational costs and easy interpretation of the results.

THE SPEAKER: Philippe Vieu is professor of Statistics at the Institut de Mathématiques of the Université Paul Sabatier in Toulouse. He has co-authored 5 books and over 60 research papers. He is associate editor of Computational Statistics and Data Analysis and of Computational Statistics. He has been an invited visitor at many universities in Spain, Europe and the US. His research interests concern, among others, non-parametric estimation, longitudinal data analysis and functional data analysis.

---

Top

Estimation under model uncertainty. Efficient inference with some bad models.

SPEAKER: Nick Longford
LANGUAGE: Anglès
PLACE: Edifici C5, Aula C5016, Campus Nord, UPC (veure mapa)
DATE: Dimarts, 9 de desembre de 2014. Hora: 12:30
ABSTRACT: Model selection has had a virtual monopoly on dealing with model uncertainty ever since models have been identified as important conduits for statistical inference. In a typical setting, a collection of models is considered and the one that achieves the best balance of the conflicting goals of adequacy (retaining all relevant parameters or covariates) and parsimony (discarding all parameters or covariates that are not important) is selected. On occasions we are aware of the chance that has intervened in the selection of a model, but we often pretend that (or act as if) the model was selected prior to data inspection and the selection was correct, and the models that happen not to have been the first choice are ignored altogether. This talk proposes an alternative based on composition, defined as linear combinations of the candidate models' estimators. Details of the general proposal are presented for ordinary regression where considerable simplification takes place. The method is illustrated on examples.

THE SPEAKER: Nick Longford is Director of SNTL, a statistics research and consulting company, and academic visitor at the Department of Economics and Business of the University of Pompeu Fabra in Barcelona. He holds a PhD in statistics from Leeds University, England, and is the author of 80 articles in peer-reviewed journals in statistics.  He has published six books, most recently 'Statistical Studies of Income, Poverty and Inequality in Europe: Computing and Graphics in R using EU-SILC', by Chapman and Hall/CRC (2014).  He has held visiting appointments and lectured at many universities all around the world.  His research interests include small-area estimation, random coefficient models, missing data analysis, causal analysis and decision theory.

---

Top

Limited-order graphical Markov models to estimate molecular regulatory networks from high-throughput genomics data

SPEAKER: Robert Castelo
LANGUAGE: English
PLACE: C5 Building, Room C5016, Campus Nord, UPC (see map)
DATE: Friday, october the 31st, 2014. Time: 12:30
ABSTRACT: Graphical Markov models (GMMs) are a powerful tool to work with families of multivariate distributions sharing a subset of conditional independence restrictions that can be represented by means of a labeled graph. Estimating the structure of GMMs from data helps investigating complex relationships between correlated random variables. This is straightforward and well-studied in the classical setting where there are more observations (n) than random variables (p), i.e., where p << n. However, technological advances in the instrumentation employed in fields like physics, engineering or molecular biology have facilitated a continuous increase of the number of objects that these instruments simultaneously observe and quantify. This leads to high-dimensional data sets where p >> n, and therefore, where classical multivariate estimation techniques and assumptions do not work anymore. In this talk, I will introduce a framework based on limited-order correlations to approach the problem of estimating the structure of GMMs from molecular data where p >> n.

THE SPEAKER: Robert Castelo is associate professor at the Department of Health and Experimencal Sciences of Universitat Pompeu Fabra (UPF), in Barcelona. He graduated in Computer Science by Universitat de Lleida (BSc) and Universitat Politècnica de Catalunya (MSc). After graduation, he spent 3 years at the National Research Institute for Mathematics and Computer Science (CWI) in Amsterdam, The Netherlands, investigating mathematical, statistical and computational aspects of graphical Markov models and got his PhD in Computer Science by the University of Utrecht. In 2002 he joined as postdoc/lecturer the UPF/CRG laboratory of Roderic Guigó working on biological sequence analysis. On 2006 he was awarded a "Ramon y Cajal" fellowship and started his own independent research at UPF where, on 2012, he became associate professor of Biostatistics and Bioinformatics.  The focus of his research over the last 8 years has been reverse-engineering the genotype-phenotype map from high-throughput genomics data, trying to embrace the complexity that follows from the coordinated action of genes by means of multivariate statistical models, such as graphical Markov models.



Top

---

Perspective Relaxation: Algorithmic methods and applications

SPEAKER: Claudio Gentile
LANGUAGE: English
PLACE: C5 Building C5, Room C5016, Campus Nord, UPC (see map)
DATE: Thursday, July the 24th, de 2014. Time: 12:30
ABSTRACT: A semicontinuous variable is a variable that can assume either the value zero or a value in a specified interval. Semicontinuous variables are modeled by a pair of variables: u and p.
The variable u is a binary variable denoting activation (u=1) or not (u=0) of the continuous variable p.  The Perspective Relaxation of a Mixed-Integer NonLinear Program with semicontinuous variables is obtained from its Continuous Relaxation by replacing each term in the (separable) objective function with its convex envelope.  Solving the Perspective Relaxation requires appropriate techniques. We discuss two well-established solution methods, the Perspective Cuts and the Second-Order Cone Programming, and two new methods based on projection of the binary variables of type u, the Projected Perspective Reformulation and its approximated version.
We present some applications (Unit Commitment, Network Design, Facility Design, CTA)  and computational results.

THE SPEAKER:
For further information on the speaker visit his webpage.

---

Top

A feasible direction interior point algorithm for nonlinear convex semidefinite programming
SPEAKER: José Herskovits
LANGUAGE: English
PLACE: C5 Building, Room C5016, Campus Nord, UPC (see map)
DATE: Monday July the 7th, 2014. Time: 12:30
ABSTRACT: The present method employs basic ideas of FDIPA [1], the Feasible Direction Interior Point Algorithm for nonlinear optimization.  It generates a descent sequence of points at the interior of the feasible set, defined by the semidefinite constraints. The algorithm performs Newton-like iterations  to solve the first order Karush-Kuhn-Tucker optimality conditions. At each iteration, two linear systems with the same coefficient matrix must be solved. The first one generates a descent direction. In the second linear system, a precisely defined perturbation in the left hand side is done and, as a consequence, a descent feasible direction is obtained. An inexact line search is then performed to ensure that the new iterate is interior and the objective is lower. A proof of global convergence of is presented. Some numerical are described. We also present the results with structural topology optimization problems employing a mathematical model based on semidefinite programming. The results suggest efficiency and high robustness of the proposed method.

[1] HERSKOVITS J . A Feasible Directions Interior Point Technique For Nonlinear
Optimization.. JOTA Journal of Optimization Theory and Applications, Londres, v. 99, n. 1, p. 121-146, 1998.

THE SPEAKER: Prof. José Herskovits Norman works on the development of Numerical Methods for Optimization and their applications in Mechanical Engineering, mainly in Structural Optimization and Stress Analysis involving variational inequalities. He is the author of a general interior point technique for nonlinear constrained optimization and a series of iterative algorithms based on this technique. These methods are employed by engineers and researchers. He also developed several methods for Structural Optimization covering a wide set of problems concerning discrete structures, such as trusses, beams and plates and also shape optimization of shells and solids. Herskovits' interior point algorithms also proved to be also strong and efficient in stress analysis of solids in contact and nonlinear limit analysis.
Click here to access his personal web page

---

Top

Screening for Osteoporosis in Postmenopausal Women: A Case Study in Interval Censored Competing Risks Data
SPEAKER: Jason Fine
LANGUAGE: English
PLACE: C5 Building, Room C5016, Campus Nord, UPC (see map)
DATE: Tuesday June the 3rd, 2014. Time: 12:30
ABSTRACT: Current US Preventive Services Task Force encourages osteoporosis screening using bone mineral density but does not specify a screening interval or ages to start and stop testing using an evidence based rationale. The current analysis explores these issues using data from the Study of Osteoporotic Fractures, the longest running cohort study of osteoporosis in the United States. Complications arise: time to osteporosis in individuals free of osteoporosis, prior fracture, and previous preventive treatment, is subject to potentially dependent censoring by fracture and preventive treatment. Endpoint definition is addressed in a competing risks framework, with a certain cumulative incidence function correctly defining the risk of osteoporosis for the screening population. The analysis of this quantity is based on intermittent bone mineral density testing. Likelihood based inference, both full and "naive", is investigated for such interval censored competing risks data, using a direct modelling strategy for the cumulative incidence functions. The screening interval is defined as a fixed time for a specified percentage of non-osteoporotic women to develop osteoporosis, accounting for the potentially dependent competing risks, which involves the use of so-called competing risks quantiles. The competing risks analysis illustrates how osteporosis risks may be precisely quantified and used to develop evidence based policy for osteoporosis screening.

THE SPEAKER: For more information on Jason Fine's research clic on http://sph.unc.edu/profiles/jason-fine/

---

Top
Vehicle routing and fleet sizing problems arising in servicing offshore oil and gas installations
SPEAKER: Irina Gribkovskaia
LANGUAGE: English
PLACE: UPC, Campus Nord. C5 Building. Room C5016 (see map)
DATE: Tuesday, March 18th 2014. TIME: 12:30
ABSTRACT: The upstream logistics in the oil and gas industry in Western Norway is a subject of ongoing research at Molde University College - Specialized University in Logistics. In this lecture we focus on fleet sizing and planning of routes and schedules for supply vessels servicing offshore installations. Supply vessels routing and scheduling is very complex and challenging due to the periodic service requirements from installations, combination of delivery and pickup services, the multi-commodity nature of demands, limited deck area at installations, time windows at supply bases and installations. An added source of complexity is weather uncertainty. We introduce several models and solution algorithms developed to handle deterministic variants of the routing problem including non-simultaneous pickups and deliveries, selective pickups, limited deck capacity, multiple commodities, periodic vessel planning and fleet composition problem, and emissions reduction through speed optimization. We also present results of several simulation studies on strategic fleet sizing with uncertainty in weather conditions, on evaluation of robustness of vessel schedules, on generation of robust and green vessel schedules, and on evaluation of different operational strategies in response to weather. Vessel routes and schedules are modelled using discrete-event simulation. Tests are performed on real data from oil and gas companies. Weather data analysis is based on historical data obtained from Norwegian Meteorological Institute for several offshore regions on the Norwegian Continental Shelf.
This is a joing work with Ellen Karoline Norlund and Yauhen Maisiuk.

THE SPEAKER:
Irina Gribkovskaia is a Professor in Quantitative Logistics with focus on operations research and a particular interest in vehicle routing and scheduling problems with applications in offshore oil and gas logistics. She received her PhD in Differential Equations from the Belorusian State University, Minsk, Belarus, where she worked 20 years at the Department of Optimization and Methods of Optimal Control. She is working at Molde University College – Specialized University in Logistics, Norway, from 1999. She is teaching courses in mathematical modeling, vehicle routing and operations management for the MSc program in Logistics. Professor Gribkovskaia publishes in EJOR, JORS, C&OR, TR, Omega etc. In recent years Professor Gribkovskaia has established close contacts within the oil and gas industry in Norway, also giving her MSc and PhD students the opportunity to benefit from this network. Students perform their research based on up-to-date problems within the industry and in close cooperation with professionals from oil and gas companies.
Related publications:

• A large neighbourhood search heuristic for a periodic supply vessel planning problem arising in offshore oil and gas operations, INFOR, 2012.
• Reducing emissions through speed optimization in supply vessel planning, TRD, 2013.
• Passenger and pilot risk minimization in offshore helicopter transportation, Omega, 2012.
• Helicopter routing in the Norwegian oil industry: Including safety concerns for passenger transport, IJPDLM, 2011.
• A simulation study of the fleet sizing problem arising in offshore anchor handling operations, EJOR, 2010.
• Lasso solution strategies for the vehicle routing problem with pickups and deliveries, EJOR, 2009.
• Routing of supply vessels to petroleum installations, IJPDLM, 2007.
• A tabu search heuristic for a routing problem arising in the servicing offshore oil and gas platforms, JORS, 2007.
• The single vehicle routing problem with deliveries and selective pickups, C &OR, 2007.
• General solutions to the single vehicle routing problem with pickups and deliveries, EJOR, 2007.

---

Top

Optimizing Personalized Predictions using Joint Models

SPEAKER: Dimitris Rizopoulos

LANGUAGE: English
PLACE: Edifici C5, Room C5016, Campus Nord, UPC (see map)
DATE: Monday, January 20th 2014. Time: 12:30
ABSTRACT: The joint modeling of longitudinal and time-to-event data is an active area of statistics research that has received a lot of attention in recent years. More recently, a new and attractive application of this type of model has been to obtain individualized predictions of survival probabilities and/or of future longitudinal responses. The advantageous feature of these predictions is that they are dynamically updated as extra longitudinal responses are collected for the subjects of interest, providing real time risk assessment using all recorded information. The aim of this paper is two-fold. First, to highlight the importance of modeling the association structure between the longitudinal and event time responses that can greatly influence the derived predictions, and second, to illustrate how we can improve the accuracy of the derived predictions by suitably combining joint models with different association structures. The second goal is achieved using Bayesian model averaging, which, in this setting, has the very intriguing feature that the model weights are not fixed but they are rather subject- and time-dependent, implying that at different follow-up times predictions for the same subject may be based on different models.

THE SPEAKER: Dimitris Rizopoulos is Assistant Professor at the Department of Biostatistics of the Erasmus University Medical Center in the Netherlands. He received an M.Sc. in statistics (2003) from the Athens University of Economics and Business, and a Ph.D. in biostatistics (2008) from the Katholieke Universiteit Leuven. Dr. Rizopoulos wrote his dissertation, as well as a number of methodological articles on various aspects of joint models for longitudinal and time-to-event data, and he is the author of the freely available R package JM that can fit several types of these models. He is also the author of the first book published in this type of models. He currently serves as an Associate Editor for Biometrics and Biostatistics, and he has been a guest editor for a special issue in joint modeling techniques in Statistical Methods in Medical Research. For more information on  his research click http://www.erasmusmc.nl/biostatistiek or http://eur.academia.edu/DimitrisRizopoulos

---

Top