ERC FUNDED PROJECTS

Making accurate predictions is a crucial factor in many systems (such as in modelling energy consumption, power load forecasting, traffic networks, process industry, environmental modelling, biomedicine, brain-machine interfaces) for cost savings, efficiency, health, safety and organizational purposes. In this proposal we aim at realizing a new generation of more advanced black-box modelling techniques for estimating predictive models from measured data. We will study different optimization modelling frameworks in order to obtain improved black-box modelling approaches. This will be done by specifying models through constrained optimization problems by studying different candidate core models (parametric models, support vector machines and kernel methods) together with additional sets of constraints and regularization mechanisms. Different candidate mathematical frameworks will be considered with models that possess primal and (Lagrange) dual model representations, functional analysis in reproducing kernel Hilbert spaces, operator splitting and optimization in Banach spaces. Several aspects that are relevant to black-box models will be studied including incorporation of prior knowledge, structured dynamical systems, tensorial data representations, interpretability and sparsity, and general purpose optimization algorithms. The methods should be suitable for handling larger data sets and high dimensional input spaces. The final goal is also to realize a next generation software tool (including symbolic generation of models and handling different supervised and unsupervised learning tasks, static and dynamic systems) that can be generically applied to data from different application areas. The proposal A-DATADRIVE-B aims at getting end-users connected to the more advanced methods through a user-friendly data-driven black-box modelling tool. The methods and tool will be tested in connection to several real-life applications.

2 485 800 €

Start date: 2012-04-01, End date: 2017-03-31

The amazing rate of progress in the computer technologies and telecommunications presents many new challenges for Optimization Theory. New problems are usually very big in size, very special in structure and possibly have a distributed data support. This makes them unsolvable by the standard optimization methods. In these situations, old theoretical models, based on the hidden Black-Box information, cannot work. New theoretical and algorithmic solutions are urgently needed. In this project we will concentrate on development of fast optimization methods for problems of big and very big size. All the new methods will be endowed with provable efficiency guarantees for large classes of optimization problems, arising in practical applications. Our main tool is the acceleration technique developed for the standard Black-Box methods as applied to smooth convex functions. However, we will have to adapt it to deal with different situations. The first line of development will be based on the smoothing technique as applied to a non-smooth functions. We propose to substantially extend this approach to generate approximate solutions in relative scale. The second line of research will be related to applying acceleration techniques to the second-order methods minimizing functions with sparse Hessians. Finally, we aim to develop fast gradient methods for huge-scale problems. The size of these problems is so big that even the usual vector operations are extremely expensive. Thus, we propose to develop new methods with sublinear iteration costs. In our approach, the main source for achieving improvements will be the proper use of problem structure. Our overall aim is to be able to solve in a routine way many important problems, which currently look unsolvable. Moreover, the theoretical development of Convex Optimization will reach the state, when there is no gap between theory and practice: the theoretically most efficient methods will definitely outperform any homebred heuristics.

2 090 038 €

Start date: 2018-09-01, End date: 2023-08-31

Our affective and motivational state is important for our decisions, actions and quality of life. Many pathological conditions affect this state. For example, addictive drugs are hyperactivating the reward system and trigger a strong motivation for continued drug intake, whereas many somatic and psychiatric diseases lead to an aversive state, characterized by loss of motivation. I will study specific neural circuits and mechanisms underlying reward and aversion, and how pathological signaling in these systems can trigger relapse in drug addiction. Given the important role of the dopaminergic neurons in the midbrain for many aspects of reward signaling, I will study how synaptic plasticity in these cells, and in their target neurons in the striatum, contribute to relapse in drug seeking. I will also study the circuits underlying aversion. Little is known about these circuits, but my hypothesis is that an important component of aversion is signaled by a specific neuronal population in the brainstem parabrachial nucleus, projecting to the central amygdala. We will test this hypothesis and also determine how this aversion circuit contributes to the persistence of addiction and to relapse. To dissect this complicated system, I am developing new genetic methods for manipulating and visualizing specific functional circuits in the mouse brain. My unique combination of state-of-the-art competence in transgenics and cutting edge knowledge in the anatomy and functional organization of the circuits behind reward and aversion should allow me to decode these systems, linking discrete circuits to behavior. Collectively, the results will indicate how signals encoding aversion and reward are integrated to control addictive behavior and they may identify novel avenues for treatment of drug addiction as well as aversion-related symptoms affecting patients with chronic inflammatory conditions and cancer.

1 500 000 €

Start date: 2010-10-01, End date: 2015-09-30

Space exploration is one of boldest and most exciting endeavors that humanity has undertaken, and it holds enormous promise for the future. Our next challenges for the spatial conquest include bringing back samples to Earth by means of robotic missions and continuing the manned exploration program, which aims at sending human beings to Mars and bring them home safely. Inaccurate prediction of the heat-flux to the surface of the spacecraft heat shield can be fatal for the crew or the success of a robotic mission. This quantity is estimated during the design phase. An accurate prediction is a particularly complex task, regarding modelling of the following phenomena that are potential “mission killers:” 1) Radiation of the plasma in the shock layer, 2) Complex surface chemistry on the thermal protection material, 3) Flow transition from laminar to turbulent. Our poor understanding of the coupled mechanisms of radiation, ablation, and transition leads to the difficulties in flux prediction. To avoid failure and ensure safety of the astronauts and payload, engineers resort to “safety factors” to determine the thickness of the heat shield, at the expense of the mass of embarked payload. Thinking out of the box and basic research are thus necessary for advancements of the models that will better define the environment and requirements for the design and safe operation of tomorrow’s space vehicles and planetary probes for the manned space exploration. The three basic ingredients for predictive science are: 1) Physico-chemical models, 2) Computational methods, 3) Experimental data. We propose to follow a complementary approach for prediction. The proposed research aims at: “Integrating new advanced physico-chemical models and computational methods, based on a multidisciplinary approach developed together with physicists, chemists, and applied mathematicians, to create a top-notch multiphysics and multiscale numerical platform for simulations of planetary atmosphere entries, crucial to the new challenges of the manned space exploration program. Experimental data will also be used for validation, following state-of-the-art uncertainty quantification methods.”

1 494 892 €

Start date: 2010-09-01, End date: 2015-08-31