Listing 1 - 7 of 7 |
Sort by
|
Choose an application
Choose an application
NRB is an IT services company based in Li`ege. A part of their business revolves around the management of changes in programs and infrastructures at the request of their clients. Any applied change may potentially provoke unforeseen incidents. To avoid this, each change is linked with a risk level that packs the probability and the level of impact of potential incidents. This way, limited resources can be adequately allocated to sensitive changes. Risk levels are assessed by human, based on several factors. However, those assessments are sometimes inaccurate, and lead to incidents or bad resources allocation. Thus, this thesis is an attempt at automating the assessment of risks levels. This problem reduces to a supervised classification problem that may be linked to the field of defects prediction with contextual parameters. Before conducting any experiment, the data is explored in details. Then, to classify changes, we try several types of algorithms, including boosting ensembles, support vector machines, neural networks, ... Transfer learning for NLP is also used with Google’s BERT. The skewness of the labels leads to the exploration of imbalanced learning solutions. This includes resampling, classes weighting and models selection (like boosting). To assess the performances of the classifiers we develop a custom metric that is inspired from the field of cost sensitive learning. A comparison with human performances shows the inability of any developed classifier to correctly detect classes. To improve this classifier, We advocate to explore in more details the field of imbalanced learning. We also advocate to use cross-validation techniques applied to temporal data. The development of a custom language model learned on a corpus of texts from NRB may also improve the performances.
Choose an application
Nowadays, the resolutions of the Traveling Salesman Problem are extremely numerous and diverse and will likely continue to spread until the famous question "P versus NP" is answered. This thesis was realized in partnership with Usitoo, a Belgian object borrowing platform based in Brussels, and thus proposes a solution in a real case application. The goal of this work is to plan deliveries and retrievals of objects in the event that certain constraints have been relaxed. The most delicate constraint that was removed is the day of the journey, which considerably increases the number of possibilities to evaluate compared to a conventional Traveling Salesman Problem. The cost representing a journey was thought so that this metric best meets the environmental issue that we are trying to address with this thesis. The suggested solution involves both exhaustive search and genetic algorithms in order to respond to the request correctly but also quickly. The final proposed algorithm was implemented in Python. In this way, we hope to respond to the request of Usitoo through a Django website presenting the obtained results. With it, they will likely be able to continue to follow their eco-responsible vocation while remaining efficient.
Choose an application
Automation is happening in all aspects of our daily lives. This work aims at improving the efficiency of NRB Service Desk thanks to machine learning and natural language processing by automating the routing of incidents tickets. This is mainly done by analysing the ticket textual content. Machine learning classifiers is compared after putting the data into the right shape. The logistic regression performed the best followed by SVM. In the end, a short study with deep learning is carried.
Machine learning --- Natural language processing --- python --- automation --- service desk --- Ingénierie, informatique & technologie > Sciences informatiques
Choose an application
Nowadays, state-of-the-art algorithms for face recognition achieve great results, even over human performances on most known testing datasets. But these algoriithms tend to be biased as the training and testing dataset are usually over represented by people sharing common facial features and color skin. Recent studies show that results on balanced dataset or representing the world population distribution tend to give lower performances and discrepancy between groups of people with different skin colors than the over-represented one. In this work, we define three metrics to evaluate this discrepancy and present three methods to reduce this discrepancy and improve results on balanced testing datasets.
Face recognition --- Machine Learning --- Deep Learning --- Fairness --- Ingénierie, informatique & technologie > Sciences informatiques
Choose an application
Renewable energies are challenging to forecast due to their intermittence. However, it is crucial for the energy transition to predict accurately what is going to be produced at different temporal resolution (short, mid or long term) to integrate them in the network. In this work, we investigate the short term horizon. We work in the practical setting of the day-ahead forecast for wind farms. The aim of this work is twofold: to help the transmission system operator (TSO) in its task of balancing the network and the market participants of the day-ahead spot market. Both tasks require to know what is going to be produced for the next day. In this work, we will try new Artificial Intelligence (i.e. AI) models for wind energy forecasting. We explore state-of-the-art Machine Learning and Deep Learning models like Random Forest, Extra Trees, Recurrent Neural Network (i.e. RNN) and Transformers. We also investigate new RNN cells (e.g. BRC, nBRC and hybrid). We create original architectures of RNNs and Transformers. To compare the models and assess the results, we use two datasets: the ORES and the Gefcom2014 dataset. The first dataset is built from ORES recording productions of wind farms located in Belgium and weather data produced by the MAR (Modèle Atmosphérique Régional) developed at the University of Liège. The second dataset is often used in the scientific community. Then, we perform a deep analysis of the results given by the best models on both datasets. Additionally, we provide perspectives of improvement and we discuss other interesting techniques to investigate further.
Deep Learning --- Machine Learning --- Energy --- Artificial Intelligence --- Energy Forecasting --- TSO --- Day-ahead forecast --- Renewable Energy --- Ingénierie, informatique & technologie > Sciences informatiques
Choose an application
The topic of this thesis is to study and compare local ranking methods applied to ensembles of trees with the aim to interpret each prediction of a model. This work contributes to the understanding of the ensembles of trees through the study of methods that locally identify the variables that are important for a prediction. The second contribution of the thesis is the further application of local ranking methods to the Gene Regulatory Network inference problem. The results of the local methods are compared with ground-truth data inferred from a simulator and their performances are compared to the state of the art method of that field.
Listing 1 - 7 of 7 |
Sort by
|