Eric Medvet runs and supervises two research labs: the Machine Learning lab and the recent Evolutionary Robotics and Artificial Life lab. In the latter, a particular focus is reserved to find new evolutionary optimization and learning techniques applied to auomatic design of autonomous robotic agents. In particular, the Voxel-based Soft Robot (VSR), a robot made of many soft blocks. VSR are intrinsically modular and suitable to be optimized in terms of shape, sensor equipment and controlling. Modularity implies, potentially , multiple configurations and robustness, caratteristics that make VSR an interesting solution for many practical scenarios. An important project, partially sponsored by Google within the prestigious Google Faculty Research Award program, is about to start and it aims to develop new learning, representations and optimization techniques to exploit the potential of VSR. Indeed , a position for a doctorate about this topic will be officialised soon.

Supercomputing, model reduction and automatic learning.

We will show recent advances to bring super computing to real life problems thanks to parametric (real time) computing by model order reduction. Data and automatic learning are playing a more and more important role in setting problems dealing with industrial and/or biomedical applications.

Several opportunities for internship and thesis both at industrial and research enviroments will be presented. Students more interested in industrial enviroment will be introduced in the activities performed at eXact lab srl, a CNR/IOM spin-off company; students attracted more by research enviroment will know about data science activities in the institute at Area SciencePark I am presently coordinating.

This talk will introduce the activities of the Artificial Intelligence and Cyber-Physical Systems Lab, and our research lines. We broadly work on three main areas: simulation, design and control of emergent behaviors of complex (cyber-physical) systems, and safe machine learning, combining tools from stochastic modelling and simulation, machine learning, and classic AI - mostly logic.

In the simulation area, we explore efficient simulation algorithms for complex systems, including dynamics on networks and of populations, using machine learning to dramatically speed up classic simulation algorithms. In this spirit, we also started recently a collaboration with Geophysicists on ML-based solutions of direct and inverse problems in this field.

In the CPS area, we work with (spatio)-temporal logic to formalize emergent behaviors of complex systems and use machine learning (Bayesian Optimization, Gaussian Processes, Deep Neural Networks and Generative Models) to provide efficient monitor, design and control algorithms. We are adding in the loop Reinforcement Learning, both as a tool for design and control, and to enforce safety constraints in RL (learning policies that are guaranteed to satisfy certain constraints).

Finally, we also work in the area of safe machine learning, at the moment mostly in the area of Bayesian-based defences again adversarial attacks.

This short talk provides an overview of the Operations Research Lab activities, which are mostly funded by the European Commission through Horizon 2020 projects. Operations Research is the discipline that “brings intellectual rigour to the decision-making process” by formulating mathematical or simulation models for optimisation. The main application area in past and current years is the air traffic management where, based on real European traffic data, we have developed optimisation models for a better use of scarce resources, such as the airspace and airport capacity. We conclude by introducing the newly funded H2020 BEACON project, which aims at modelling decisions made in the presence of incomplete or uncertain information, and with limited rationality, by agent-based computational economics.

My research is focused on the development of efficient procedures able to connect techniques of Machine Learning and Formal Verification for the analysis of large-scale spatially-distributed complex systems for which standard model checking procedures are not feasible. Interdisciplinary research has the potential to bridge the limit of the single methodology, exploiting the skills of the other. Formal verifications, as computational proof methods, can help in better understand, learn, and control the behaviors of such systems. On the other hand, machine learning methodologies, efficient, and easy to use, can be exploited to scale formal analysis. In this talk, I will give a brief overview of how logic frameworks can be exploited to describe (language specification), check (verification algorithms), and synthesize (learning techniques) these complex systems. Furthermore, I will discuss some open problems and ongoing works that may be attractive to students.

Accelerating high performance computing algorithms with Artificial Intelligence

Scientific computing is a broad field, and its applications are often computationally intensive. Several aspects of modern algorithms used in scientific computing can be accelerated and guided through artificial intelligence. In this short presentation I will discuss some of the applications we are currently working on, and how artificial intelligence could be used to accelerate them.

Evolutionary computation is the idea that (optimization) problems can be solved by mimicking the evolutionary processes that happen in nature. I will present one of the most prominent methods in evolutionary computation: genetic programming, where entire programs evolve. I will also present the recent advancements in the field and some interesting open problems, with the opportunities to work on them.

I will be introducing my main fields of interest for my research, focused on both applied and theoretical statistics. Precisely, I recently worked on: applications of hierarchical Bayesian models to political science, epidemiology, sports; mixture models; specification of robust prior distributions in Bayesian inference. I am an R package developer and I am used to work within the R environment. Possible (and broad) thesis topics: predictive sports models, Bayesian theory, clustering.