Rommert Dekker

Big Data in Shipping

Quite recently, ships are obliged to send information on their status (position, speed) to a general platform. Next public access to this so-called AIS has been ensured, such that everybody can see which ships are around (except from small ships). One would expect that the availability of such an amount of real-time data would generate a wealth of applications, but todate there are only few applications of that data. In this talk we will analyse the role of these data in facilitating applications and which – essential –data is lacking so far. We will first focus on arrivals and departures of ships in ports, next we will consider the analysis of ship delay and their use in timetables and finally we will present some AIS applications to date.

Alexander Goldenshluger

Statistical inference for the M/G/infinity queue

The subject of  this talk is statistical inference on the service time distribution and its functionals in the M/G/infinity queue. In particular, we will discuss three different observation schemes with incomplete data on the queue: observations of arrivals and departures without identification of customers,
observations of the superposed arrival-departure point process and observations of the queue-length (number-of-busy-servers) process.  In these settings we derive some probabilistic results on the processes involved and construct estimators of the service time distribution with provable accuracy guarantees. The problems of estimating the service time expectation and the arrival rate are discussed as well.  We will present also some results on comparison of different estimators.

PRESENTATION

Moshe Haviv

Queueing paradoxes

The talk will survey three paradoxes and one anti-paradox emerging in queueing systems. The paradoxes are Breass, Downs-Thomson and Javons. The anti-paradox is that the other line is not that short after all.

PRESENTATION

Nathan Kallus

Dynamic Assortment Personalization in High Dimensions

We demonstrate the importance of structural priors for effective, efficient large scale dynamic assortment personalization. Assortment personalization is the problem of choosing a best assortment of products, ads, or other offerings (items) to specifically target a particular individual or consumer segment (type). This is a central problem in revenue management for e-commerce, online advertising, and multi-location brick-and-mortar retail, where both types and items number in the thousands to millions. Efficient use of data is critical in this large-scale setting, as the number of interactions with customers is limited – definitely not in the trillions – so it is infeasible to learn each one's preferences independently. Furthermore, learning preferences is not enough: the goal of personalization is revenue, not mere knowledge. In dynamic assortment personalization, the retailer chooses assortments to learn preferences and optimize revenue simultaneously.
We formulate the dynamic assortment personalization problem as a discrete-contextual bandit with m contexts, where type (individual or segment identity) is the context and assortments of n items are the arms. We assume that each type's preferences follow a simple parametric model with n parameters (multinomial logit); in all, there are mn parameters. Existing literature suggests that, over T interactions, order optimal regret is mnlog(T), which is infeasibly large in most e-commerce settings.
In this paper, we impose natural structure – a small latent dimension for the parameter space, or low rank on the matrix of mn parameters – on the problem. This structure reduces the number of parameters to estimate and the regret incurred in estimating them. In the static setting, we show that this model can be efficiently learned from surprisingly few interactions. We propose an efficient algorithm to learn the model that converges globally whenever the model is learnable. In the dynamic setting, we show that structure-aware dynamic assortment personalization can have regret that is an order of magnitude smaller than structure-ignorant approaches.

Michael Katehakis

Simple Data Driven Policies for MDPs

Markov decision processes (MDPs) have a variety of applications, not just the classical OR fields but also in other directions such as computer science,  engineering and biology/medical science. We first give a brief survey of the state of the art of the area of computing optimal data driven (adaptive) policies for MDPs  with unknown  rewards and or transition probabilities. 
Then, we present three simple algorithms for adaptively optimizing the average reward in an unknown irreducible MDP. The first algorithm uses estimates for the MDP and and chooses actions by maximizing an inflation of the estimated right hand side of the average reward optimality equations. The second is based on estimating the optimal rates at which actions should be taken and the third is based on generalized Thompson sampling ideas.  For the first we show that the total expected reward obtained by this algorithm up to time n is within O(ln n) of the reward of the optimal policy, and in fact achieves asymptotically minimal regret.  Various computational challenges and simplifications are discussed. 
(joint work with Wesley Cowan, PhD, Rutgers University) 

Ger Koole

Data analysis and validation of call center models

Queueing models are used for scheduling purposes in many queueuing models, but they are never validated.
We analyzed data from a Dutch call center and validated some common features such us Poisson arrivals and exponential talk times.
We also compared the overall performance with models of different levels of precision in modelling.
(joint work with Sihan Ding, April Li, Rob vd Mei and Raik Stolletz)

PRESENTATION

Dimitrios Mavroeidis

Predictive Maintenance

Predictive Maintenance algorithms aim to identify early signs of deteriorating equipment condition allowing for timely scheduling of maintenance visits thus preventing unplanned downtime and customer inconveniences. The importance of this problem for several industries has led to the development of various predictive maintenance techniques in the fields of statistics, operations research and machine learning.
A big obstacle commonly faced by researchers is related to the limited availability of real-world use-cases and data to validate the existing techniques and to develop novel approaches. In this talk we will present a predictive maintenance use-case related to Philips medical devices, highlighting the specific use-case challenges as well as the solutions that have been developed.

Nazanin Nooraee

S-curve Prediction Models for High-Frequent Wind Turbine Data

Developments of technologies support collection and storage of huge amounts of data. Given this data, scientists can investigate what information is contained and try to explain what has happened or predict what will happen in the future. However, drawing inference and making reliable conclusions, even from large data sets, often require advanced statistical models.
In this talk, we discuss and compare various parametric and non-parametric statistical models for an industrial application with high-frequent data. More precisely, we were interested in modeling the relation between wind turbine power and wind speed. In order to model the wind turbine power curve, which has an S-curve shape, we applied different kinds of 5-parameter logistic functions and smoothing functions on one season of data collected at a frequency of every ten minutes. The mathematical models were evaluated with the mean square error and they were compared to a theoretical minimum. Additionally, the model prediction were validated on different time windows.
(joint work with Stella Kapostria and Sándor Kolumbán (TU/e))

Ohad Perry

Service Systems with Dependent Service and Patience Times

Most queueing models for service systems in the literature are assumed to have independent primitive processes (arrival, service, abandonment, etc.). However, data shows that the patience of a customer may depend on that customer’s service requirement. In this talk we study the impacts that such a dependency has on key performance measures (waiting times, queue length, proportion of abandonment and throughput), and on optimal capacity decisions. In particular, we consider a system with a single pool of many statistically-homogeneous agents serving one class of statistically-identical customers whose service requirements and patience times are dependent. Since the assumed dependence structure renders exact analysis intractable, we propose a fluid approximation which is characterized via the entire joint distribution of the service and patience times. To evaluate the effects of the dependence, we employ bivariate dependence orders and copulas, and provide structural results which facilitate revenue optimization when a staffing cost is incurred. Simulation experiments demonstrate that our fluid approximation is accurate and effective.
(joint work with Allen Wu and Achal Bassamboo, Northwestern University)

Georg Pflug

Decision making under uncertainty: data-driven modeling

As we understand it, data-driven modeling means that a minimum of assumptions (or no assumptions) are made a-priori and we let the observed data "speak for themselves". To put it differently, the the data-driven approach should be a nonparametric one.
In this talk, we consider a multistage stochastic decision model as it appears in portfolio optimization, option pricing, hydrostorage management, energy trading etc. The uncertain parameters of such a model form a (multivariate) stochastic process, for which typically a parametric form is assumed (in the simplest case such as in the Black-Scholes formula just with one single parameter).
For the numerical calculation of optimal decisions, a discretization (a scenario tree or a scenario lattice) is needed and can be found by approximation techniques. We review briefly these techniques, but put the emphasis on nonparametric scenario generation. While large data sets are of course desirable for the model quality, the technique can also handle quite small data sets. The same nonparametric technique allows also to reduce large scenario trees in size with keeping much of their stochastic properties as we demonstrate with examples.
In addition, we consider the problem of model misspecification. By using a nonparametric notion of model distance, we are able to define decision strategies, which are robust against small model misspecifications. A relation between the size of the data set and the model ambiguity can be given by recent inequalities due to Villani and others. An illustrative example from hydrostorage management is given.

PRESENTATION

George Shanthikumar

A Framework for Prescriptive Empirical Operations Management

We provide a framework for prescriptive empirical modeling with specific attention to overcoming structural and statistical errors. This is achieved through operational statistics and objective operational learning which are built on the basis of data integration and cross validation. We will illustrate how regularization in sample approximation approaches and data driven robust optimization with cross validation relates to operational objectives and operational statistics. We will also illustrate how data driven modeling using data mining and econometric modeling with machine learning can be used here.
(joint work with F. Qi)

Kalyan Talluri

Facility Location Decisions from Public Data

There has been a tremendous increase in the size and availability of public data, yet it is not clear how a firm might put it to good use. Analytical studies rarely seem to go beyond summary statistics and attractive visualizations. In this paper we present an application based only on publicly available data in which a restaurant chain makes a facility-location decision: where to locate a new restaurant, of what type, and in which price range. We combine Yelp review data sets, with demographic, geographic and restaurant inspection data to build a model of demand and use it to formulate an optimization problem that recommends the top $k$ locations.
(joint work with Muge Tekin (Universitat Pompeu Fabra, Barcelona))

Evrim Ursavas

Locating LNG Bunkering Stations

The growing awareness of the environment and new regulations of the International Maritime Organization and the European Union has forced ship-owners to reduce pollution. Liquified natural gas (LNG) is one of
the most promising options for accomplishing this reduction for inland waterways and short sea shipping. However, an extensive LNG infrastructure has not yet been developed due to low commercial demand and
demand is still low due to the absence of LNG fuel facilities. To overcome this problem, refuelling facilities need to be located in strategic locations allowing for alternative investment structures. To this end, we
develop mathematical models that determine locations for refuelling stations where terminal-to-ship and truck-to-ship bunkering alternatives are analyzed.We consider the characteristics of an LNG network, such as the boil-off eff ect (from loading and storage) and dynamic demand by accounting for factors like capacity. We consider cases where capacity expansion and reduction is beneficial.We develop a lower bound heuristic based on lagrangian relaxation technique. We derive bounds by employing a heuristic to obtain feasible solutions of the lagrangian solution. We perform our experiments for the waterway network in the Arnhem-Nijmegen region in the Netherlands. Extensive experiments are provided to study the possible future scenarios.

Jelle de Vries

Determinants of Safe and Productive Driving: Empirical Evidence from Long-haul Cargo Transport

Using GPS data of 370 long-haul trips in India, survey data of 49 drivers, and ERP data, this study examines the role of driver personality characteristics in predicting risky and productive driving. The results show that more conscientious drivers display more risky driving behavior. More extravert drivers are less productive, whereas driver safety consciousness positively relates to productivity. These results can serve as a starting point for further studies into how long-haul transport companies may use individual driver characteristics in their training and selection procedures to meet operational safety and productivity objectives.
(joint worh with René de Koster, Debjit Roy, Serge Rijsdijk)

PRESENTATION

Spyros Zoumpoulis

Customizing Marketing Decisions Using Field Experiments

We investigate how firms can use the results of field experiments to optimize promotion assignments between customers. We evaluate seven widely-used segmentation methods using a series of two large scale field experiments. The first field experiment is used to generate a common pool of training data for each of the seven methods. We then validate the seven optimized policies provided by each method together with uniform benchmark policies in a second field experiment. The findings reveal that model-driven methods (Lasso regression and Finite Mixture Models) performed the best. Some distance-driven methods also performed well (particularly $k$-Nearest Neighbors). However, the classification methods we tested performed relatively poorly. The precision of the data varied with the level of aggregation, and the model-driven methods made the best use of the information when it was more precise. The model-driven methods also performed best in parts of the parameter space that are well represented in the training data. The relative performance of the methods is robust to modest differences in the settings used to create the training and validation data. We explain the poor performance of the classification methods in our setting and describe how it could be improved.

PRESENTATION