DOCTEUR INGÉNIEUR CIVIL
specialized in Mathematical Programming (identification, classification, optimization & control)
Name: Dr. Ir. Frank Vanden Berghen
Place & Date of birth: Brussels, 1974-12-31
|Home:|| 11 Chemin des 2 Villers
7812 Ath (V.N.D.) - Belgium
Tel.: +32 (0) 479 99.27.68
Ph. D. in the field of mathematical engineering with a specialization in continuous optimization at the faculty of polytechnics at the University of Brussels (ULB - Université Libre de Bruxelles), Belgium (obtained with the highest distinction)
Thesis : CONDOR: a constrained, non-linear, derivative-free parallel optimizer for continuous, high computing load, noisy objective functions.
degree of "Diplôme d’Étude Approfondie (DEA) en Sciences Appliquées" at the faculty of polytechnics at the University of Brussels (ULB - Université Libre de Bruxelles), Belgium
Final year project : I realized an unconstrained optimizer for continuous, high computing-load, objective function when the derivatives are not available.
degree of "Ingénieur civil" specialised in computer science at the faculty of polytechnics at the University of Brussels (ULB).
Final year project: I realized a multi-input multi-ouput fuzzy direct auto-adaptative controller with multiple step ahead prediction (MSDAFC). This controller was used for the FAMIMO european research project.
Secondary studies, Athénée Royal Ath, Belgium
|May 2007- now||
Founder of the “TIMi” company (Machine Learning – Big Data - Business Intelligence)
I created the TIMi company. The objective of this company is to provide state-of-the-art analytics and advanced analytics tool for big data, predictive modeling, AI and Business-Intelligence. The tools from TIMi technically outperforms any other off-the-shelf commercially available predictive datamining tool for a fraction of their price. Please see the “Awards” section to know more about the tools from TIMi. To summarize: We are the best data scientists of Europe!
|May 2007- now||
Development of a machine learning software for the TIMi company (Machine Learning – Big Data - Business Intelligence)
The “TIMi Suite” was cited inside the very elitist report “Advanced Analytic Quadrant 2016” from the Gartner company. This places the “TIMi Suite” as a world-level leader in the Data Science field on the same level as other prestigious multinational companies such as: Oracle, Tibco, Mathworks, Salford system, etc.
I am the lead coder for the development of these tools. I was also the project leader for a large part of the development of the applications, supervising a team of 10 people. I designed the algorithms and the architecture of all the softwares.
|May 2007- now||
Development of a Big Data software for the TIMi company (Machine Learning – Big Data - Business Intelligence)
The Big Data tool is named “Anatella”. It’s one component of TIMi. Anatella has been compared to many other ETL tools and is currently the fastest and the most scalable of all the Big Data tools available on the market. For example, it’s several orders of magnitude faster than Spark and much more scalable and reliable. Anatella is also one of the only self-service ETL tool available on the market (everything is done with the mouse). This places TIMi at the forefront of the international Big Data/Analytical field.
|May 2007- now|| |
Development of a various “Big Data” and “Machine Learning” solutions for many private companies using TIMi as the analytical platform (Machine Learning – Big Data - Business Intelligence)
|August 2008- July 2009||
Development of a GPS application (like TOMTOM) for Punch Telematix
I was in charge of the “Street Indexing” module and the “Road Snapping” module. I created from scratch a small embedded-DB engine where the street data are indexed using a structure named “PatTrie” that is similar to the “Bit Index” of the “SyBase IQ” database. This highly optimized indexing structure allows constant-time fuzzy match on the “street names” on the DB (independently of the size of the DB).
|February 2008- August 2008||
Development of a scientific application for analysis of “Small Light Scattering” for the Rheology Laboratory of the KUL
The application includes FFT transformation and eigenvalues computation of large images (6000x6000). It was done inside the Qt4 framework.
July 2007 -
Consultant at Business&Decision Belgium (machine learning)
Research engineer in the Analysis of Large Social Networks (machine learning)
I was working for the VADIS company to develop a software that analyses large social networks based on mobile phone communication logs. The analysis includes segmentation in clusters (or tribes) and node ranking inside the clusters. The softwares are based on I/O efficient external memory algorithms to be able to process large networks(2000 GB). The objective is to extract from the network some indexes that could be use for marketing/profiling purposes.
Research engineer in robust optimal control of batch processes (control)
I worked for the IPCOS company to realize a software that computes the optimal set of parameters of a hybrid PID controller (including terms for feedback, feedforward and filtering) for a strongly non-linear process. The parameters are computed taking into account the full non-linearity of the batch process along the setpoint trajectory. No linearity assumption is made at any point. The final result is higher precison in the solution parameters. The parameters are computed using a special version of my optimizer (CONDOR) that is able to work with non-differentiable non-linear constraints. The constraints include different kind of stability margins (mixed L_2/H_nfity control). On a first industrial computer benchmark (produced by UCB), CONDOR found an optimal set of parameters that reduces the tracking error to 0.1% of the currently implemented industrial solution while maintaining the same stability margins.
Research engineer in KDD (Machine Learning - Knowledge Discovery in Databases)
I worked for the WegenerDM/VADIS company. I realized a datamining/profiling application able to process vast amount of data (several Gigabytes) in a few minutes. The typical size of the tables that are analyzed is more than 10^8 lines and 10^5 columns. This tools is now used everyday to develop large direct marketing campaigns for a number of prestigious companies (Citibank, Procter&Gamble, YvesRocher,...). This tool has been compared with other large scale statistical analysis packages (SAS, KXEN, NORKOM, SPSS) and gives higher quality results thanks to advanced coding and advanced mathematics. The computation time of this tool is a fraction (1/1000) of the computation time of other solutions.
Research engineer in optimization of Dry Seals Designs (optimization)
I worked shortly at the Computational Department of the BURGMANN Industries in Wolfratshausen, Germany. Burgmann produces currently nearly all high-technology seals in the international market. The european community wants to change this situation and open the seal market place to other companies. One way of reaching this goal is to publish a set of "standard seals designs" that can be used in a vast amount of applications. Once these designs are defined, many industries will be able to compete for the production of the seals. The final result is the opening of the seal market. The technical specifications of these "standard seals" must be carefully chosen because they will be used by all manufacturers in Europe. The optimal specifications of the "standard seals" were computed using my optimizer, CONDOR. The seal optimization process involves an objective function based on a seal simulation. This simulation includes thermodynamic and plastic deformations of the seal combined with the fluid dynamics simulation of the gas in the leakage. The objective function is thus highly non-linear, high-computing-load and undefined on many points of the input space.
Research engineer in optimization algorithms for continuous functions.(optimization)
I worked on the METHOD LTR European project with the university of Florence . The goal of this project is to optimise the shape of turbo-machines. We run a simulation of the gas flow inside the turbine. Based on this simulation, we compute the quality of the shape. We iterate on different shapes until we find the best shape. My job was to write the optimisation code that chooses what's the next shape to try. I use trust region method optimization on multivariate Lagrange interpolator techniques. The final result is my optimizer, CONDOR.
Research engineer in Fuzzy control (machine learning, non-linear identification & control)
I worked at the development of the FAMIMO ESPRIT LTR European project in collaboration with other European universities and SIEMENS Automotive. The research activities focus on two different areas. The development of new approaches in the area of fuzzy identification and control , and the implementation of a Matlab toolbox that will integrate all the techniques developed by the partners taking part in the project (see http://iridia.ulb.ac.be/~famimo/).
Research engineer in Real-Time image classification (machine learning)
I worked on the GLAVERBEL project: a real-time classifier for the classification of glass defects (world première). This classifier is implemented as a modified TFTP server. It receives images from the TCP/IP network, computes the class, sends back the results to an Oracle Database and to the factory computer. Then, the factory computer decides how to cut the glass (commercial optimisation). The percentage of good classification is 94%.
Development of a pre-press application for the publishing industry.
I created a large java application called "Advertedge". This software processes postscript and PDF files to include gray marks on the edges of the pages. The final objective is to obtain a drawing on three sides of the book when it is closed. This application is currently used all over the world in the biggest printing factory (casterman, donelley,...).
Development of a time series prediction tool (machine learning)
This tool is currently used by D'IETEREN, one of the biggest Belgian car reseller for sales prediction. This tool is based on Lazy Learning techniques.
Frank Vanden Berghen, CONDOR, a parallel, direct, constrained optimizer for high-computing-load, black box objective functions Proceeedings of the "Third MIT conference on Computational Fluid and Solid Mechanics", Elsevier, june 14-17, 2005.
Frank Vanden Berghen, CONDOR: a constrained, non-linear, derivative-free parallel optimizer for continuous, high computing load, noisy objective functions. PhD Thesis, University of Brussels (ULB - Université Libre de Bruxelles), June 2004, Belgium
Frank Vanden Berghen, Hugues Bersini CONDOR, a new parallel, constrained extension of Powell's UOBYQA algorithm: Experimental results and comparison with the DFO algorithm Journal of Computational and Applied Mathematics, Elsevier, Volume 181, Issue 1, 1 September 2005, Pages 157-175 (also available electronically on the Sciences Direct website)
Hussain Aziz Saleh, Frank Vanden Berghen Human genome behaviour: a powerful mechanism for optimizing the use of space technology in surveying networks design GPS Solutions, Springer-Verlag GmbH, Volume 9, Number 3, September 2005, Pages: 201 - 211
Frank Vanden Berghen Design et implémentation d’un nouvel algorithme d’optimisation continue non-linéaire dans le cas sans contrainte Specialisation Research Project presented to obtain the Diploma of Deeper Study in Applied Sciences (DEA), University of Brussels (ULB - Université Libre de Bruxelles), June 2003, Belgium
Simone Pazzi, Francesco Martelli, Vittorio Michelassi, Marco Giachi, Frank Vanden Berghen, Hugues Bersini, Intelligent Performance CFD optimization of a centrifugal impeller accepted to the Fifth European Conference on Turbomachinery, March 2003, Prague (CZ).
Frank Vanden Berghen A tutorial on Q-learning algorithms internal IRIDIA technical report
Frank Vanden Berghen Hugues Bersini, Régulation directe adaptative et predictive sur plusieurs pas de temps pour processus à plusieurs entrées et plusieurs sorties chapter of a book entitled « Commande Floue I - de la stabilisation à la supervision » published by Hermes-Science -Lavoisier edition
Frank Vanden Berghen, Edy Bertolissi, Antoine Duchâteau, Hugues Bersini, Direct Adaptive Fuzzy Control for MIMO Processes, Accepted to the FUZZ-IEEE 2000 conference, San Antonio, Texas, 7-10 May, 200
Frank Vanden Berghen, Edy Bertolissi, Antoine Duchâteau, Hugues Bersini Régulation directe adaptative et prédictive sur plusieurs pas de temps pour la commande floue de processus à plusieurs entrées et plusieurs sorties, internal IRIDIA technical report
Frank Vanden Berghen Développement d’un régulateur flou à plusieurs entrées et plusieurs sorties adaptatif et prédictif sur plusieurs unités de temps-Utilisation et évaluation des performances de ce régulateur Final Research Project presented to obtain the degree of "ingénieur civil" specialised in computer science.
|August 19-20, 2016||
Big Data Analytics Summit (Lima, Peru)
As a guest speaker, I am giving a technical presentation on TIMi (algorithmic, technical architecture, past and future evolutions). TIMi possesses a unique set of features (automated modeling, self-service data analytics, etc.) and I explain why this places TIMi apart from all other solutions currently available and why it revolutionizes the way in which companies are managing their big data and predictive analytics initiatives.
The organizers from the conference selected a very limited set of prestigious speakers, such as Gilberto Titericz Jr, that is amongst the top Data Scientist in the world following the analytic Kaggle competitions.
|October 23, 2015||
TIMi, Biggs & IBM event (Bogota, Colombia)
This event regrouped nearly all the data scientists from Colombia. I participated to an open discussion with the Head of IBM for Latin America where we talked about the future of advanced analytics.
|September 21, 2015||
Data Innovation Lab Event (Brussels, Belgium)
The title of the event was “Real Life Advanced Analytics examples with TIMi”. I demonstrated some live cases with TIMi. Other consultants gave nice example of usage of TIMi: PwC presented a SNA (Social Network Analytics) example, AXA presented a churn model for car insurance, Agilytic.be presented the work they did for VOO/BeTV on cross-sell models.
|June 19-21, 2008||
Computing and Statistics (ERCIM ’08) (Neuchâtel, Switzerland)
I gave a presentation on the efficient implementation with the LARS-lasso algorithm and talked about some experimentation.
|July 18-22, 2005||
22nd IFIP TC 7 Conference on System Modeling and Optimization (Turin, Italy)
I gave a presentation on the CONDOR optimizer: the algorithmic details and the numerical results.
|June 14-17, 2005||
Third MIT international Conference on Computational Fluid an Solid Mechanics (Massachusetts Institute of Technology, Cambridge, USA)
I gave a presentation on the CONDOR optimizer: the algorithmic details and the numerical results.
|May 29-30, 2000||
international BELGIUM/FUZZY 2 conference (Faculté des Sciences Appliquées, Mons, Belgium)
I have presented some results that we obtained at IRIDIA while we were working on the FAMIMO ESPRIT LTR European project.
Following the benchmarking results obtained on the universally recognized TPC-H benchmark, Anatella is now the fastest and the most scalable data management tool (ETL tool) on the planet.
Selected amongst the top 100 most innovate software companies at the international/world level by Red Herring international
We went to Los Angeles to get our prize! More details here (look at “business-insight”).
Selected amongst the top 100 most innovate software companies in Europe by Red Herring international
We went to Amsterdam to get our prize!
1st place at the European Data Innovation Hub: Euroclear Hackaton
The aim is to identify key fields in unstructured financial legal documentation. Data consists in thousands of legal documents. We won the first place in this contest with a solution that extracts the “maximum trading amount” out of these legal documents with an accuracy over 97%. The whole project duration was less than 2 days. More details here: https://timi.eu/customer-stories/banking/text-mining, here and here.
Céline Theeuws obtained the 9th place on a total of 415 participants for a modeling competition named “Construction d’un score d’appétence en vente croisée pour un produit d’assurance lors d’une campagne télémarketing”. Céline used TIMi to create all the predictive models.
Kaggle Competition: AXA Driver Telematics Analysis
Dr. Ir. Colin Molter obtained the 10th place on a total of 1528 participants for a modeling competition. Colin used TIMi to create all the predictive models.
Selected to represent Belgium during the Royal Mission in the presence of Princess Astrid to Colombia and Peru.
We represented Belgium activities linked to big data, advanced analytics and predictive modeling solutions.
TIMi was elected 6th best Predictive Analytic Solution by a team of expert data scientists.
|June and January 2014|
Innov Iris Prices
In June 2014, Vadis won the “InnovIrisDay price” from the Brussels Region because of RANK, a predictive modeling software that I developed for them (alone) a few years ago. TIMi is in every way superior to RANK. In January 2014, Reaktor wins the “InnovIris price RISE” thanks to a solution based on TIMi.
This is a world-wide datamining competition. The objective is to predict the rating (number of "stars" given to a movie: from 1 to 5) that a specific individual will give to a specific movie. The goal of the AusDM 2009 Analytic Challenge was to encourage the discovery of new algorithms for ensembling or 'blending' sets of expert predictions. I obtained the 5th place on a total of 30 participants. I used the “TIM software suite” for datamining for the competition.
KDD cup 2009
“KDD cups” (KDD stands for “Knowledge Discovery from Database”) are the most famous competitions in the datamining field. In 2009, the competition had 2 different challenges: I obtained the ranking 22 at the first challenge (also called the “large dataset” challenge because the dataset had 15000 columns) and I obtained the ranking 18 on the second challenge (also called the “small dataset” challenge because the dataset had a “normal-size” of 300 columns). There were more than 1200 participant. This ranking is based on the average quality (AUC) of 3 predictive models that are built using data coming from the "Orange" company (it's the number 1 of the French Telecom). The 3 predictive models to build are a Churn model, a Upselling Model and a "Propensity-to-buy" (appetency) model. I used the “TIM software suite” for datamining for the competition.
Pacific Asia KDD cup 2007
This is a world-wide datamining competition. The objective is to create a predictive model for a cross-selling application : “credit card” to “Mortage”. I obtained the 6th place on a total of 47 participants. I used an ancestor of the “TIM software suite” for datamining for the competition.
"Young fellowship Award for exemplary research in Computation Mechanic" from the MIT (Massachusetts Institute of Technology)
The CONDOR optimizer was recognized as a valuable advance in the field of continuous optimization. As a result, it was installed and made publicly available inside the world-famous NEOS Server. See the webpage: http://neos.mcs.anl.gov/neos/solvers/ndo:condor/AMPL.html
The CONDOR optimizer was also included inside the famous Decision Tree for Optimization Software hosted at:
This places CONDOR among the top optimization softwares and expands the field of my research on optimization at the international level.
|Secondary school||- 1992 :||
1st price in the Informatics Olympic Games in Belgium. I was send with 2 students to represent Belgium at the World Informatics Olympic Games where I classified among the top 50 out 200.
|- 1991 :||
5th price Expo-Sciences, Namur, Belgium (subject: calculation of hidden faces on 3D models)
|- 1989 :||
1st price at Expo-Shell (subject: binary representation & assembler programming )
French: mother tongue
Dutch: secondary school knowledge. I also obtained the international certification "Nederlands als Vreemde Taal - graad 2" in the following categories : oral and written comprehension, oral and written expression (There are only 3 grades).
English: good level (spoken and written fluently).
I am involved in the teaching activities of the company “TIMi Americas”. Amongst different teaching assignments, “TIMi Americas” is giving machine learning and Big data courses in these contexts:
I am giving Datamining and predictive Analytics courses to train the data scientists from various private companies on how to create highly-efficient predictive models for various types of applications (mainly churn, cross-selling and up-selling models). For example:
I created YouTube tutorial videos to learn how to do various ETL tasks with Anatella, the high-performance Big data/ETL tool included in TIMi.
I was involved in the teaching activities of the IRIDIA laboratory. I gave laboratories of C++ and java in the faculty of engineering. I also gave advanced java formation for enterprises (see www.technofutur3.be ).
MS SQL-Server, MySQL, Informix, (superficial knowledge in Oracle, Teradata)
I possess a deep knowledge in everything related to Big Data & Advanced Analytics. I am familiar with 99% of any classification/regression/clustering algorithm for machine learning and AI. During my thesis, I used extensively interpolation/regression techniques, optimization techniques (continuous and discrete), fuzzy control and the classical optimal robust control theory, high efficiency linear algebra routines. I had the opportunity to use extensively the lapack and blas libraries, self-adaptive fuzzy controllers, decision trees, lazy learning, multivariate lagrange interpolators, neural networks, evolutive and gradient-based optimization algorithms.
low level TCP-IP socket programming in C&Java ,...; CGI ; client-server applications;...During my work at IRIDIA, I had the opportunity to construct and use a Beowulf of 8 nodes.
I am familiar with Visual Studio .NET, the TIMi suite, R-Studio, Anaconda, MySQL, Php, JS, Oracle, Informix, Subversion, Git, 3Dstudio Max 2, Photoshop, Apache, SAS, Microsoft Office.
Unix (Linux and Solaris) and Windows 98/XP/Vista/8/10 are more than familiar to me.
My objective is to allow any company or person to extract knowledge out of their data. I firmly believe that we can live in a better world if we start using more the data that surrounds us.
To help people to efficiently use their data, I created TIMi. TIMi is a data analysis platform to make machine learning, big data & AI. TIMi is totally revolutionary: 1000 times faster, 1000 times more scalable, virtually no need for any heavy hardware infrastructure, 100% in self-service and operated entirely with the mouse.
My ambition with TIMi is to create the analytical platform of tomorrow. This pushes me to develop new algorithms to create more efficient, more precise and faster software solutions. To reach these goals, TIMi is using a combination of high-level advanced mathematics and optimized low-level programming (C&ASM).
Sport: squash, karting , fencing, scuba diving, KiteSurf, gliding, badminton, Horse riding,..
Hobbies: programming, SF books, comics, cinema, rock'n roll dancing, RPG,...