Description
Date depot: 20 mars 2019
Titre: Scalable and Automated Nonparametric Analytics Titre de la these: Analyse non paramétrique automatique et scalable
Directeur de thèse:
Pietro MICHIARDI (Eurecom)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
In the past 50 years, there has been an
increasing demand for a principled approach to the maintenance of industrial
equipment. However, maintenance processes are typically triggered by either
malfunctioning, or by a simple time-based maintenance schedule. The development
of techniques designed to help determine the condition of in-service equipment
in order to predict when maintenance should be performed, is a recent effort
that goes under the name of predictive maintenance.
In the context of this PhD Thesis, we
will explore novel approaches to design, analyse and validate interpretable
machine learning models, that not only output predictions (e.g. whether a
system is in a normal, functioning state or not) but also which are the rules
that determine such predictions.
In particular, we will work on Boolean
function approximation, in that their structure readily accommodates expressing
rules in normal forms, such a disjunctive normal forms. The main challenges we
will address are related to: 1) the discrete nature of the modelling space,
which calls for optimisation heuristics or relaxation techniques for loss
minimisation; 2) algorithmic scalability in face of very large training sets;
3) automatic support in decision making, through the use of the reinforcement
learning paradigm.
This Thesis is supported by SAP, which will
provide its operational data, and will provide a safe “sand box environment” to
evaluate the contributions of this work
Doctorant.e: Mita Graziano