Communication Engineering

ANTONELLO RIZZI
ANTONELLO RIZZI
ENRICO DE SANTIS
ENRICO DE SANTIS

Introduction to Machine Learning and data driven modelling. Soft Computing, Computational Intelligence. Basic data driven modelling problems: clustering, classification, unsupervised modelling, function approximation, prediction. Generalization capability. Deduction and induction.
Induction inference principle over normed spaces. Models and training algorithms. Distance measures and basic preprocessing procedures.
Optimization problems. Optimality conditions. Linear regression. LSE and RLSE algorithms. Numerical optimization
algorithms: steepest descent and Newton’s method.
Fuzzy logic principles. Fuzzy induction inference principle. Fuzzy Rules.
Classification systems: performance and sensitivity measures. K-NN Classification rule.
The biological neuron and the central nervous system.
Perceptron. Feedforward networks: Multi-layer perceptron. Error Back Propagation algorithm. Support Vector
Machines. Automatic modeling systems. Structural parameter sensitivity. Constructive and pruning algorithms.
Generalization capability optimization: cross-validation and Ockham's razor criterion based techniques.
Min-Max neurofuzzy classifiers; standard and regularized training algorithm. ARC, PARC; Principal Component Analysis; Generalized Min-Max neurofuzzy networks. GPARC.
Swarm Intelligence. Evolutionary Computation. Genetic algorithms. Particle Swarm Optimization, Ant Colony
Optimization. Automatic feature selection.
Fuzzy reasoning. Generalized modus ponens; FIS; fuzzyfication and e defuzzyfication. ANFIS. Basic and advanced
training algorithms: clustering in the joint input-output space, hyperplane clustering.
Outline of prediction and cross-prediction problems: embedding based on genetic algorithms.
Applications and case studies: micro-grids energy flows modelling and control, Smart Grids optimization and control,
classification of TCP/IP traffic flows.
Mining of frequent patterns and rule extraction in large data bases (Big Data Analytics).

KNOWLEDGE AND UNDERSTANDING. The basic principles of the design of automatic systems for machine learning are provided (classification problems, clustering, functional approximation and prediction) based on Computational Intelligence techniques (neural networks, fuzzy logic, evolutionary algorithms). Students who pass the final exam will be able to read and understand texts and articles on advanced topics in the field of Soft Computing and Computational Intelligence (Neural networks, optimization meta-heuristics, fuzzy systems).

APPLICATION CAPABILITIES. Students who pass the final exam will be able to apply the methodological principles and algorithms studied for the design of innovative machine learning systems, in multidisciplinary contexts.

AUTONOMY OF JUDGMENT. Students who pass the final exam will be able to analyze the design requirements and choose the machine learning system that best suits the case study.

COMMUNICATION SKILLS. Students who pass the final exam will be able to compile a technical report and create an appropriate presentation aimed at documenting any design, development and performance measurement work concerning a machine learning system.

ABILITY TO LEARN. Students who pass the final exam will be able to independently continue deepening the topics covered in class, carrying out the necessary continuous learning process that characterizes professionalism in the ICT field.

Elementary notions of Geometry, Algebra, Differential Calculus, Signal Theory, Information Theory, Informatics, Digital Signal Processing.

First part:
Introduction to Machine Learning and data driven modelling. Soft Computing, Computational Intelligence. Basic data driven modelling problems: clustering, classification, unsupervised modelling, function approximation, prediction. Generalization capability. Deduction and induction.
Induction inference principle over normed spaces. Models and training algorithms. Distance measures and basic preprocessing procedures.
Optimization problems. Optimality conditions. Linear regression. LSE algorithm. Numerical optimization
algorithms: steepest descent and Newton’s method.
Fuzzy logic principles. Fuzzy induction inference principle. Fuzzy Rules.
Classification systems: performance and sensitivity measures. K-NN Classification rule.
The biological neuron and the central nervous system.
Perceptron. Feedforward networks: Multi-layer perceptron. Error Back Propagation algorithm. Support Vector Machines. Automatic modeling systems. Structural parameter sensitivity. Constructive and pruning algorithms.
Swarm Intelligence. Evolutionary Computation. Genetic algorithms. Particle Swarm Optimization, Ant Colony Optimization. Automatic feature selection.
Fuzzy reasoning. Generalized modus ponens; FIS; fuzzyfication and e defuzzyfication. ANFIS. Basic and advanced training algorithms: clustering in the joint input-output space, hyperplane clustering.
Outline of prediction and cross-prediction problems: embedding based on genetic algorithms.

Second part:
The second part of the course begins by exploring performance measures, which are fundamental for evaluating the effectiveness of machine learning algorithms. It addresses the problem of overfitting, discussing measures of structural complexity and the Occam's Razor criterion. The importance of cross-validation, particularly N-fold cross-validation, is a focal point to ensure the generalization of models. Additionally, specific topics related to classification and functional approximation are covered, such as the costs of misclassification and the challenges posed by unbalanced datasets.
The course then delves deeper into Support Vector Machines (SVM), a powerful tool for supervised learning used both for classification and regression. The basic principles of SVMs are explored, such as the definition of support vectors and the optimization of the margin through an optimal hyperplane. The mathematical formulation is discussed along with the importance of the "kernel trick," which allows SVMs to tackle nonlinear classification problems. Solving dual problems through optimization techniques and practical considerations for implementing SVMs are integral parts of the module.
In the context of Smart Grids and energy management, a specific module introduces Computational Intelligence techniques for energy system management. Topics such as energy transition, microgrids, and renewable energies are discussed. Energy Management Systems (EMS), storage systems, and their modeling, and Renewable Energy Communities (REC) are explored, applying Computational Intelligence techniques to optimize energy flow management. Particularly, Fuzzy Systems optimized with Evolutionary Algorithms are discussed. The use of Dynamic Programming and multi-objective optimization is illustrated, with a focus on advanced prediction systems like LSTM networks that are useful in forming effective prediction modules for EMS.
Deep learning occupies a significant portion of the course, with an overview of the fundamental principles and the differences between deep learning, machine learning, and artificial intelligence. Various neural architectures are analyzed, including Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN), with particular emphasis on LSTMs for managing sequential data. The main libraries for synthesizing deep learning models are explored, with a focus on TensorFlow. The concept of computational lattice and gradient calculation of complex computational structures through Automatic Differentiation techniques is discussed. A specific practical session demonstrates how to synthesize machine learning and deep learning algorithms using TensorFlow.
The concept of prediction and forecasting using traditional and deep learning techniques is introduced. The importance of prediction in the context of energy system optimization is highlighted.
Finally, the course addresses the problem of predictive maintenance, exploring techniques and models to predict failures and optimize maintenance in industrial sectors. The importance of applying machine learning techniques is discussed, and case studies are analyzed to illustrate best practices and challenges in implementing these technologies.

Kruse, R., Borgelt, C., Braune, C., Mostaghim, S., & Steinbrecher, M. (2016). Computational intelligence: a methodological introduction. Springer.

Lecture notes and slides available at teacher's site

The course is organized as a series of lectures and case study illustrations.

It is strongly recommended to attend classroom lessons.

The final exam consists in the evaluation of a homework. The topic of the homework is usually agreed with the teacher.

Not applicable