Moncef Garouani

Université Toulouse Capitole, France

Abstract : Machine learning (ML) has penetrated all aspects of the modern life, and brought more convenience and satisfaction for variables of interest. However, building such solutions is a time consuming and challenging process that requires highly technical expertise. This certainly engages many more people, not necessarily experts, to perform analytics tasks. While the selection and the parametrization of ML models require tedious episodes of trial and error. Additionally, domain experts often lack the expertise to apply advanced analytics. Consequently, it necessitates frequent consultations with data scientists; nevertheless, such collaborations tend to cost the delays, which can lead to risks such as human-resource bottlenecks. As the complexity of these tasks increases, so does the demand for support solutions. In response, the field of automated ML (AutoML) is a data mining-based formalism that aims to reduce human effort and speedup the development cycle through automation. It can be applied to create pipelines of traditional ML models and ensembles, or to search for neural network architectures (NAS). In this regard, existing approaches include Bayesian optimization, evolutionary algorithms as well as reinforcement learning. These approaches have focused on providing user assistance by automating parts or the entire data analysis process, but without being concerned on its impact on the analysis. The goal has generally been focused on the performance factors, thus leaving aside other important and even crucial aspects such as computational complexity, confidence and transparency. In the talk we will present an overview of the main components of AutoML, the search spaces representing ML models, the search strategies, and the Explainability of automated AI.

Franck Dufrenois

LISIC- IMAP team Université du Littoral Cote d'Opale

Abstract : (kernel) Fisher Discriminant Analysis
Fisher Discriminant Analysis (FDA) and its kernel version, Kernel FDA, are common supervised learning methods for dimensionality reduction and class separation. The objective of this tutorial is to introduce you and expand your knowledge of this technique. First, I will recall the notions of projection and reconstruction. Then, I will present the main ingredients of FDA : the betweenclass, the within-class and total scatter matrices. Next, I will introduce the FDA’s principle and the corresponding optimization problem to solve it. The dimensionality of the projective subspace will also be discussed. I will also discuss on the main drawback of FDA, i.e the possible singularity of the scatter and how to prevent from it. FDA is based on the assumption that data are linearly separable. In a second part, I will present how the « kernel trick » allowed to generalize LDA to nonlinear data sets. We will see how the FDA criterion is formulated and solved in this case. Some computational aspects and memory constraints will be also covered. Finally, I will end the talk by some simulations on different datasets.

Octavian Curea

ESTIA, Bidart, France

Abstract: The utilization of Cyber-Physical Systems (CPS) and Artificial Intelligence (AI) in building energy management increase the energy efficiency and sustainability in our working or living spaces. These technologies have brought advancements in how we monitor, control, and optimize the energy consumption. Cyber-Physical Systems are the core of this transformation, integrating computational, communication, and control components into physical systems, creating interconnected environments. In the context of buildings, CPS allows real-time data collection from sensors distributed throughout the structure and the control of the demand. The data provided by the sensors includes temperature, light, air quality, occupancy, energy consumption and production of the distributed energy sources as PV or small WT. Artificial Intelligence plays an important role in analyzing the available data to optimize energy management. Machine learning algorithms are employed to forecast energy demand, identify potential wastage, and suggest real-time adjustments to save energy. Some of the benefits of the integration of Cyber-Physical Systems and Artificial Intelligence in building energy management are the improvement of the occupants’ comfort, the reduction of the energy costs and carbon footprint, the promotion of the sustainability.

Mohammed Ramdani

Faculty of Science and Techniques of Mohammedia, University Hassan II, Casablanca, Morocco

Mostafa Bellafkih

Institut National des Postes et Télécommunications | INPT, Rabat, Morocco

Abstract:

Noura Yousfi

Laboratoire d’Analyse, Modélisation et Simulation Département de Mathématiques et d’Informatique Faculté des Sciences Ben M'Sick, University Hassan II, Casablanca, Morocco

Abstract: Optimization in computing science
Optimize and optimization are derived from Latin Optimus, which means 'best'. The term to optimize was used in the mid-19th century with the meaning of making the best or the most of something. In the 21st century, it has been extensively utilized in technical fields, such as economics, calculus, mathematics, engineering, neuronal networks, and so forth. Optimization is a fundamental concept in the field of computing science. Its primary objective is to make the system's efficiency while using limited resources. It can occur at several levels : In optimization problems, scheduling and transportation are two application cases that are often encountered; Optimization algorithms aim to find the best solution that minimizes or maximizes a particular objective function; The application program developers is to design their programs in such a way that they make the most use of limited and expensive resources; Also using programming languages, which are algebraic modeling languages makes coding easier and shorter. Many tasks learning can be formulated as optimization problems, and these can be improved by modern algorithms. Gradient descent and stochastic gradient descent are by far the most popular differential optimization strategy used in machine learning and deep learning. It is used when training data models, can be combined with every algorithm and is easy to understand and implement. On the other hand, gradient-free optimization methods are increasingly successful in many fields of engineering or medicine where the functions to be optimized are black box type, very expensive to evaluate, and defined on discrete/continuous mixed spaces. The purpose of this tutorial is to provide a brief overview of the gradient descent algorithm, its current usage, and its advantages and disadvantages. In addition, it presents the most prominent gradient-free optimization methods that have been developed in recent years, particularly in genetic algorithms.

Mohamed Youssfi

Professor Researcher in Department of Mathematics and Computer Science, ENSET, Hassan II University

Abstract: Mohamed Youssfi received his PhD (Doctorat de 3ème cycle ) from The Mohammed V University of Rabat, Morocco in 1996 in Computer Science and his PhD ( Doctorat d’Etat ) from The Mohammed V University of Rabat, Morocco in 2015 in Parallel and Distributed Systems. He has served as Professor Researcher in the Department of Mathematics and Computer Science, ENSET, Hassan II University of Casablanca since 1996. His scientific publications are focused in Computational Intelligence, Parallel and Distributed Systems, Multi Agent Systems, Web Semantic, High Performance Computing, Big data, and Distributed Artificial Intelligence. He has (co)authored more than 100 scientific research articles in indexed journals, refereed conferences, edited volumes and books.

Gilles Régnier

PIMM Laboratory, Arts et Métiers, CNRS, CNAM, France. Arts et Métiers Rabat campus

The computational costs of injection molding simulations have been increasing in the past years due to the higher complexity of the embedded models. This is especially problematic in case of using such simulation models for optimization routines or sensitivity analyses. One way to overcome this challenge is by having a surrogate model, also known as a metamodel, of these high-fidelity simulations, which provides a cheaper way to perform these types of analyses. These surrogate models can play an important role in the case of the injection molding of semi-crystalline polymers to model the flow-induced crystallization process. To date, most commercial software do not explicitly take polymer crystallization into account leading to various errors in the fill predictions as well as the calculation of warpage and shrinkage. This is mainly due to the common complexity of the models used to describe crystallization and the challenging respective model parameter identification process under injection molding conditions. To close this gap, in this thesis, the feasibility of using surrogate modeling to identify modeling parameters is first studied. This is then followed by the implementation of a thermo-mechanical crystallization model in order to describe the flow-induced and quiescent crystallization of a semi-crystalline thermoplastic material a during injection molding. The crystallization model is defined alongside crystallization-dependent viscosity, PVT and solidification models in the commercial software Autodesk Moldflow Insight 2021 using the Solver API feature. The model parameters are identified using a calibration scheme that employs three surrogate models representing the simulated pressure results to perform a multi-objective optimization. The fill predictions as well as the calculated pressure fields are presented using the calibrated model parameters in comparison to those measured during the actual injection molding of a polyoxymethylene part with different process conditions. The results show major improvements in the predictions of the pressure signals as well as the filling status of the produced parts and the estimated skin layer thicknesses formed under high-shear conditions. Additionally, the calibrated models are tested using various mold geometries to assess the calibrated models' performance.