2017-11-24
| Speaker: Andrés Giovanny Pérez Coronado Affiliation: PhD student in Mathematical Engineering, Universidad EAFIT Title: Forecasting of the crime with matching between murders and Google trends Abstract: Models of crime forecasting could use internet data, to understand how
people make decisions about our own security, looking for information on
search engines. These queries on internet describe how a topic like
“murders” is important in the people daily lives. So, this presentation
brings together police data from the database called “SIEDCO plus” and
Google trends from 2015 to 2017, using the Kernel Density Estimation
(KDE) to analyzing and to matching information.
We try to understand how the statistical data of police has relations to forecast crimes with people
queries. The police information specifies geographic coordinates
of murders, quantity, date, kind, centroid of police quadrant (division
on field of policing). This aims to build a paper on probabilistic
model to forecast of organized crime, linked
with social media.
Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m.
|
2017-11-23
| Speaker: Santiago Lopez Restrepo Affiliation: Research assistant, Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT
Title: Ensemble-based data assimilation to improve Chemical Transport Models
Abstract: Ensemble-based data assimilation, which refers to the sequential use of the direct measurements to create accurate initial conditions for model runs, is one of the most commonly used approaches for real-time forecasting problems. In each assimilation step, a forecast from the previous model simulation is used as a first guess, using the available observation this forecast is modified in better agreement with these observations. The ensemble Kalman filter technique (EnKF), as the most known ensemblebased data assimilation technique, is used to assimilate in situ, satellite and aircraft measurements from different applications of CTM around the world . In addition, EnKF is chosen because it can be easily combined with covariance localization . An accurate covariance localization is essential to reduce spurious covariances during assimilating.
Slides: N/A Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-11-03
| Speaker: Andrés Yarce Botero Affiliation: Research assistant, Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT Title: The Variational approach on the road to Data Assimilation (DA) for Chemical Transport Models (CTM) Abstract: DA is the process whereby observations are incorporated into the model state of a numerical model of that system. In our case, we use the model LOTOS-EUROS (LE), a Chemical Transport Model (CTM) used to predict the atmospheric spatio-temporal concentration of several gasses and their deposition rates. Two approaches for the solution of DA problems were introduced before at the last seminar session: The Statistical (Ensemble Based Kalman Filter) and the Classical (Variational) method. This talk will focus on the Variational approach, illustrating the development of this technique and preparing the ground for the exploration and implementation of two major Variational DA algorithms. Two recent Ph.D. theses from TU Delft 2016 (Fu, 2016; and Lu, 2016), are reviewed to present and help to understand the different perspectives of the Variational methods that have been studied before tackling the state of the art on the subject developed in the Applied Mathematics group of Professor Heemink. The Variational approach for DA of existing data (e.g., satellite observations), integrated with LE, will help for us construct a methodology that will be used for estimating the impacts on natural ecosystems from the deposition of air contaminants. * Lu, S. (2016). Variational data assimilation of satellite observations to estimate volcanic ash emission (doctoral dissertation). Technische Universiteit Delft, Delft, The Netherlands. * Fu, G. (2016). Improving volcanic ash forecast with ensembled-based data assimilation (doctoral dissertation). Technische Universiteit Delft, Delft, The Netherlands. Slides: [ pdf ] Video: [ url ] Place: 27-304 Time: 11 a.m. - 12 m. |
2017-10-27
| Speaker: David Barrera Affiliation: Postdoc, École Polytechnique, Palaiseau, France Title: Least Squares Regression for Non-Stationary Designs* Abstract: The main goal of this talk is to present a series of new results concerning the rate of convergence for least square estimates in the case in which the sampling process (the "design") is not i.i.d. Our results are given in the setting of nonparametric regression, and they cover the corresponding estimates in the i.i.d. case (as given,
for instance, in [1]) without any essential loss in the respective rates of convergence nor the introduction of additional hypotheses in order to carry out the proofs. They justify also a more general -but very natural- interpretation of the least-squares regression function as the “best” approximation to the response function in a given statistical experiment, and provide further theoretical ground for the research on numerical methods in which a non-stationary evolution has to be considered.
We illustrate these results and their aforementioned interpretation in the numerical context by looking at estimation problems in which the i.i.d. setting is either not satisfiable or not convenient, emphasizing in particular the Markovian setting. We also illustrate the relevance of these tools in the error analysis of Monte Carlo algorithms like the one in [2].
[1] Györfi, L.; Kohler,M.; Krzyzak, A. and Walk, H. (2002). A Distribution-Free Theory of Nonparametric Regression. Springer Ser. Statist. [2] Fort, G.; Gobet, E. and Moulines, E. (2017) MCMC Design-Based Non-Parametric Regression for Rare Event. Application for Nested Risk Computations. To appear in Monte Carlo Methods Appl.
* This talk was first given during the 11th International Conference on Monte Carlo Methods and Applications, held at HEC Montréal (Canada) on the days July 3-7, 2017.
Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m.
|
2017-10-25 | Speaker: Raúl Ramos Pollán Affiliation: * Professor / Researcher at Universidad Industrial de Santander (UIS), Bucaramanga, Colombia * Leader Large Scale Data Analytics, Center for Supercomputing at UIS * Project Manager, GALILEO Information Center for Latin America Title: Using convolutional networks in practice Abstract: Convolutional networks are currently producing state of the art
results in many image analysis tasks. Designing and training such
networks is no trivial task both in terms of network architecture and
computing resources, and in many occasions one is obliged
to resort to fine tuning existing pretrained networks. This seminar
shows strategies for using convolutional networks in three different
scenarios: (1) to build models when image datasets are small, (2) to
learn features of the data to improve the interpretability
of the models and (3) to exploit simulated data to augment the training
process of the networks. Application domains showcased will be medical
imaging, astrophysics and image semantics.
Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 3 p.m. - 4 p.m. |
2017-10-20
| Speaker: Juan David Palacio Domínguez Affiliation: PhD student in Mathematical Engineering, Research Group in Functional Analysis and Applications, Universidad EAFIT
Title: Towards a general framework for the Repositioning Problem in Bicycle-sharing Systems
Abstract: The Repositioning Problem in Bicycle-sharing Systems (BSS) aims to find a
set of routes able to serve all the system stations by picking up or
delivering bicycles according to a previous demand estimation and a
fixed number of available bicycles. To address
this kind of problems, we design a general framework for vehicle
routing problems with several features. So far, we have included pick up
and deliveries, single and multiple vehicles, and heterogeneous fleet.
From the solution strategies perspective, this
framework includes a new extension of our mixed-integer formulation now
able to deal with
multiple vehicles.
We also include simple heuristics and some meta heuristics based on
greedy, variable neighborhood search and path relinking algorithms for
the single vehicle case.
Slides: [ pdf ] Video: N/A
Place: 01-924 Time: 11 a.m. - 12 m. |
2017-10-13 | Speaker: Leandro Fabio Ariza Jiménez Affiliation: PhD student in Mathematical Engineering, Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT Title: A network analysis approach for metagenomic binning Abstract: Metagenomics is the application of modern genomics techniques to the study of communities of microbial organisms directly in their natural environments. In metagenomics, the assignment of genomic fragments to the corresponding taxonomic group, e.g. species, genera or higher taxonomic groups, is commonly referred to as “binning”, a procedure wherein each of the sequences is placed into an imaginary bin representing ideally only fragments belonging to this group. Since this is essentially a data clustering problem, here we attempt to develop and implement unsupervised strategies to address such problem. In particular, in this talk we will present the progress made by following a novel approach for the binning of genomic fragments based on similarity networks and community detection algorithms.
Slides: N/A Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-09-29 | Speaker: Juan Carlos Arango Parra Affiliation: PhD student in Mathematical Engineering, Research Group in Functional Analysis and Applications, Universidad EAFIT Title: Diffusion Kernels on Statistical Manifolds Abstract: We present a conceptual revision of the article Diffusion Kernels onStatistical Manifolds, by John Lafferty and Guy Lebanon. The main objective of the article is that when classifying empirical distributions obtained as sampling of members of a parametric family M of probability distributions, it is convenient to use the kernel of the heat equation of (M, g) where g is the metric of Fisher of the family M. In agreement with the properties of this kernel of the heat, it is transformed into a Mercer kernel that allows to classify data within the framework of support vector machine (SVM) algorithms. Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m.
|
2017-09-22
| Speaker: Diana Paola Lizarralde Bejarano Affiliation: PhD student in Mathematical Engineering, Research Group in Functional Analysis and Applications, Universidad EAFIT Title: Mathematical strategies in the study of epidemiological models based on nonlinear differential equations Abstract: In this talk, we will present a simple analysis of model parameters
which could be influenced by control strategies. Using simulations, we
determine how these strategies affect the value of Basic Reproductive
Number to evaluate the impact of the infected
population. Since there is not enough information about entomological
parameters and initial conditions of the studied models, we aim to
formulate interval-valued parameters by considering the uncertainty to
obtain robust solutions. Finally, to calculate these intervals, we formulate the inverse
problem associated, to explore different strategies proposed in the
literature to solve the original problem taking into account the
restrictions obtained from the stability analysis and the
information about the phenomenon. Slides: [ pdf ] Video: N/A
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-09-15 | Speaker: Eduardo García-Portugués Affiliation: Assistant Professor at the Departament of Statistics of the Carlos III University of Madrid Title: Smoothing-based inference with directional data Abstract: Directional data arises when measuring observations on a circumference, a
sphere, or, in general, a hypersphere. This type of data appears
naturally in several applied disciplines, gaining increasing popularity
in recent years in bioinformatics. Statistical
treatment of directional data requires careful rethinking, since the
constant-norm setting makes the direction of
a vector its only relevant piece of
information. As a consequence, dependence relations, orderings and even
standard statistical objects such as the mean suffer from non-trivial
modifications whose study generated a considerable statistical
literature in the last half-century. Nonparametric
smoothing-based inference has proven itself as a highly useful
methodology for directional data, excellently combining flexibility and
tractability. In this talk we present two recent nonparametric proposals
for the estimation of the density and regression
curves, respectively, in the context of directional data.
The first part of the talk considers the problem of estimating a
directional density under the common assumption of rotational symmetry. A
new operator that rotasymmetrizes any
directional density is introduced, which allows to construct a kernel
density estimator with directional data. The basic asymptotic properties
of the estimator are derived, bandwidth selection is discussed, and its
finite sample performance is illustrated
in a simulation study.
The second part of the talk focus on inference for the regression
function of a scalar response on a directional predictor. From a
nonparametric perspective, a new local-linear estimator for the
regression curve is presented. This estimator is used as a pilot
for assessing the goodness-of-fit of
a parametric regression model, whose asymptotic distribution is
obtained. The performance of the testing procedure
is illustrated in a simulation study and, finally, the test is applied
to check a commonly assumed hypothesis in bioinformatics. Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-08-25
| Speaker: Andrés Giovanny Pérez Coronado Affiliation: PhD student in Mathematical Engineering, Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT
Title: Foresight to dismantling criminal networks
Abstract: Guaranteeing citizen security and preventing violence from organized crime becomes more difficult every day. Criminals adapt quickly and organize themselves in complex systems for law enforcement agencies. Thus, through the use of spatial-temporal modeling this doctoral research aims to build state of the art mathematical and statistical tools that allows anticipating criminal behavior and dismantle criminal networks. Through the mix of networks and graph theory, probability theory, criminology and a unique dataset that will be build from confidential information of the Colombian Police and Social Media (twitter). I propose to create models that allow identifying the best way to break up criminal networks and predict crime reorganization. Therefore, this doctoral research stage proposes to integrate probability theory within network and graph theory using Social Media information processed through semantic analysis.
Slides: [ pdf ] Video: N/A Place: 27-304 Time: 11 a.m. - 12 m. |
2017-08-11
| Speaker: Myladis Rocio Cogollo Flórez Affiliation: Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT Title: New approach to estimation of non-normal process capability indices using artificial intelligence techniques Abstract: In industrial processes it is very common that the data associated with
different characteristics of a product are not normally distributed, so
that the application of traditional methods of estimation of process
capability indices (PCI) can lead
to erroneous results. Adaptations of traditional PCIs assuming
non-normal data, have been considered in the literature, such as:
transformation techniques, Clements percentiles method, and the Burr
percentile method. However, these techniques require the knowledge
of the distribution of the data and the use of tabulated values to
obtain the estimation of the associated parameters. Another recent
proposal is to use an artificial neural network (ANN) model to estimate
PCIs. Although it is an innovative idea in the context
of the analysis of industrial processes, the model requires the
formulation of distributional assumptions, the use of tabulated values,
and also does not satisfy a systematic procedure for the construction of
an ANN model. In this
research proposal the construction of an ANN model under a formal
construction methodology and without distributional assumptions is
proposed. In addition, this model could be used in
real time. Slides: [ pdf ] Video: N/A
Place: 27-304 Time: 2 p.m. - 3 p.m. |
2017-08-04 | Speaker: Henry Laniado Rodas Affiliation: Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT Title: Robust and Nonparametric Statistical Tools for Big Data in Neuroscience Abstract: How
the human brain works is one of the most beautiful questions we have
been asking our whole life, and it is amazing how the statistical field
can help to answer this question. Functional
Magnetic Resonance Imaging (fMRI) is one of the top techniques within
the neuroimaging field that relates with this topic. The aim of fMRI
data analysis is to determine which regions of the brain are either
activated or inactivated with respect to an experimental
design. In order to do this, one must consider a large partition of the
whole brain, consisting of a set of very small cuboid elements called
voxels, each of one representing a million of
brain cells. After the patient is subjected to some type
of stimulus (auditory, visual, mechanical), the result of the entire
procedure is an image of the brain, showing some zones that were
positively related to the experiment and the
rest of the area, represents the non-activated zones, i.e. the areas
that did not have relation at all with the experiment.
Note that they are actually clusters of voxels—perhaps
hundreds of them. This leads to the statistical problem of how to manage
this dataset to obtain an image as the explained previously. Complexity
and massive amount of this kind of
data, and the presence of different types of noises, makes the fMRI
data analysis a challenging one; that demands robust and computationally
efficient statistical analysis methods for high Dimensional data. The
classical approach is to consider in each voxel
of the brain a General Linear Model to estimate if the observed signal
is significantly similar to the expected signal, in order to decide
activation or not activation for each voxel. However, we need to be
aware of the assumptions of the models, in order
to consider the results as valid and then obtaining correct statistical
inference, but with this kind of data, these assumptions do not always
hold. So, the adopted methodology to address fMRI statistical analysis
lacks of robustness, although it is computationally
efficient. We propose here a non-parametric a robust statistical
techniques to face this problem, while maintaining efficient
computational time, comparable with the classic method. In other words,
we are interested on finding a method that provides to the
neuroimaging field with a balance between robustness and computational
efficiency. Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-07-28
| Speaker: Olga Lucía Quintero Montoya Affiliation: Research Group in Mathematical Modeling - GRIMMAT, Universidad EAFIT
Title: Sensitivity and uncertainty sources in numerical modeling to forecast atmospheric systems: High-resolution WRF model simulations in urban valleys applied to air quality issues.
Cooperative Project: Universidad de Antioquia, Universidad EAFIT Contributions of the Grupo de Ingeniería y Gestión Ambiental (GIGA), Escuela Ambiental, Facultad de Ingeniería, Universidad de Antioquia, Medellín, Colombia, Mathematical Modeling Research Group and Biodiversidad, Evolución y Conservación Research Group, Universidad EAFIT, DIAM at TUDELFT, TNO The Netherlands. Executive summary The Colombian air quality dynamics analysis will be develop through the use of the LOTOS-EUROS Model from TNO looking for the cooperation between Universities and Institutions. In a previous Cooperative project we proposed at least seven theoretical aspects related to the Data Assimilation Schemes regarding the Backtracking localization strategies for Ensemble Kalman Filter and the traj4D-Var proposed by (Fu et al, 2015 and Lu et al, 2015), also the application of Observational impact analysis algorithm (TSBOI-MM) (Verlaan and Sumihar, 2016). Another goals related to the theoretical aspects of the particular modeling and data assimilation of the Air Quality within the Colombia Domain were also adressed.
During the development of the mentioned research, we realize that the short-term meteorological forecast with the numerical model WRF (Weather Research and Forecasting model) is a very relevant stage for an integrated and accurate research on the field. We also have demonstrated the challenges and opportunities to contribute scientifically and practically due to this approach (López et al, 2017; Pinel et al, 2017; Rendón et al, 2017; Posada et al, 2017; Rodríguez et al, 2017). Also TNO experts declared this as project to research during next years, conditioned to The European conditions, not as limited as ours. Mathematical models are the main scientific tool for understanding and predicting the potential response of the atmosphere to perturbations such as different meteorological conditions, altered emissions, and land use/land cover modifications. These models contain the non linear and complex differential equations that rule the physics of the phenomenon such as Navier-Stokes equations, and they must to be solved numerically and present a series of challenges related to their sensitivity and uncertainty when solutions are required in a non homogeneous domain also under restrictions in boundary and initial conditions. The main goal of this research is to identify significant sources of uncertainty in the short-term meteorological forecast with the numerical model WRF (Weather Research and Forecasting model), reduce the uncertainty for the solution of the equations for modeling and forecast, determine and explain the sensitivity of the model and the numerical solution due to the aforementioned difficulties for its solution, perform a successful coupling with LOTOS EUROS model for Chemical and transport dynamics in pseudo real time, as well as to discuss the causes and provide scientific evidence of the implications for the air quality modeling in the Aburrá river valley. The urban population growth generates an increase in urbanization and pollutant emissions with negative consequences for the environment and public health. Further, there are several physical phenomena occurring in the low atmosphere of the cities that affect the concentration of pollutants and the chemistry composition of atmosphere over urbanized areas. These problems are more critical in complex terrains where ventilation is limited. During the last years, air pollution problems have become more frequent in the Aburrá river valley as a result of the combined effects of the complex topography, the accelerated population growth and the associated surface alterations and pollutant emissions. In this sense it is necessary to improve our capability to understand and predict local meteorological phenomena and the associated air pollution episodes, and involve this knowledge on the decision-making processes in the metropolitan area. Our aim is to predict episodes of atmospheric pollution to make efficient decisions that allow guaranteeing air quality and human health, providing conditions to the future human exposure model to pollution to be developed by BEC research group. In cooperation with Universidad de Antioquia now we propose to study the uncertainty measurements and develop a framework for uncertainty reduction in a prediction step for Weather forecast via WRF Model, taking into account its NON real time computability and computational load. We also propose the evaluation of both models WRF/OPEN LOTOS-EUROS on the characterization of daily cycle for meteorology and pollutants dispersion in Valle de Aburrá (Aburrá Valley) also the study of scenary for the Aburrá Valley and its environmental implications through WRF/OPEN LOTOS-EUROS. Coupling and static/dynamic downscaling of the models are challenging task depending on models and data availability. (Rendon et al, 2014; Posada et al, 2016). Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |
2017-07-21
| Speaker: Juan Carlos Rivera Affiliation: Research Group in Functional Analysis and Applications, Universidad EAFIT Title: Combinatorial
optimization: applications, models and solution methods
Abstract: This
presentation has as objective to present a proposed 3-years research project
presented to a call for projects from Universidad EAFIT. In this research, it
is proposed to study different solution methods for combinatorial optimization problems
with applications on vehicle routing, scheduling, timetabling, and/or finance.
Particularly, this research project proposes to design solution procedures that
can be used to deal different kind of problems. Some examples from the state of
the art will be presented. Among the methods that could be used we can find
exact methods, heuristics, metaheuristics and matheuristics methods. The last
procedures are especially important due to most of the problems related with
this research belong to the NP-Hard class. As result of this project, in addition to the solution procedures and algorithms, we hope to include real
applications, to participate in scientific events, to publish scientific
articles, and to integrate master and doctoral students.
Slides: [ pdf ] Video: [ url ]
Place: 27-304 Time: 11 a.m. - 12 m. |