Numerical experiments with global optimization for space trajectory planning, using a library developed at ESA Candidate: Tommaso Aldinucci Graduated: april 2017
Global Optimization methods for semi-supervised learning
A quasi-Newton based approach to train a semi-supervised Support Vector Machine. Candidate: Lorenzo Norcini – graduated Feb 2017
Machine Hearing
content-based feature engineering for automatic identification of musical genre Candidate: Tina Raissi
Semi-supervised training via a lagrangean approach
Implementation of global optimization algorithms, preferably in python, to trai a SVM in which some of the data has no label Many algorithms exist for S3VM – exact (branch and bound) and heuristic (global optimization). We aim at implementing some
Forecasting time series with Support Vector Regression
A comparison between classical (ARIMA) forecasting methods for time series and regression based on Support Vector Machines. A huge set of economic time series is available to train and validate foreasting methods Skills required: basic computer science skills; python might
Iterative LASSO Framework for feature selection
Feature selection can be performed by the Least Absolute Shrinkage and Selection Operator (LASSO), a regression method which imposes a penalty on the absolute value of the regression coefficients. LASSO presents some limitations since it can’t succeed in recovering the right set of features
Training Support Vector Machines using second order information
The goal of the thesis is to solve the Support Vector Regression problem through second order methods to achieve fast convergence and good accuracy. A primal, unconstrained and twice differentiable formulation of the problem is used. The Representer theorem allows
Optimization algorithms for Recurrent neural networks
Recurrent neural networks (RNNs) are know to be extremely powerful yet effectively impossible to train with standard first order methods when the training sequences exhibit long term dependencies. The aim of this thesis is to develop a novel optimization algorithm
Optimization methods applied to portfolio selection problems
In this thesis many risk-based asset allocation models are compared, focusing particularly on the Risk Parity strategy. Risk Parity is an allocation method used to build diversified portfolios that does not rely on any assumptions of expected returns, thus placing
Do not waste past computations: Learning and Optimization for Optimal Circle Packing
A simple task: use machine learning in order to avoid useless local searches to be started. Try to find new putative optimal configurations learning from past trials. Skill required: just python – most numerical experiments in circle packing have already been