New Faculty Seminars

Dec 18, 2019, 11:00 - 13:30 
Room A1.6 - "Alan Turing" building

“The importance of being positive: a system theoretic perspective”

Positive systems are dynamical systems whose trajectories staynonnegative for all future times if started from nonnegative initialconditions and forced by nonnegative inputs. Since positivity is a fun-damental property of a huge amount of physical systems, it is not sur-prising that positive models have been extensively studied in the lastdecades, highlighting a remarkable set of properties and applications.One of the strengths of linear positive systems relies in the simpli edstability analysis, which proves to be a surprisingly strong advantagein the time-delay setting. Indeed, while studying the stability of lin-ear delay systems can be a hard task, the positivity assumption vastlysimpli es the analysis, leading to easy to check necessary and sucientconditions, as recently shown in a number of insightful works. Thenice properties of positive delay systems gave us the main motivationfor a series of works, which we survey in this talk, whose key idea isproviding a method to systematically construct a positive representa-tion of an arbitrary delay system, in order to apply to the latter thestrong properties that only hold for its positive representation. Thishas led to novel delay-independent stability conditions for some gen-eral classes of linear delay systems. We briey review these resultsfor delay di erential, delay di erence and coupled delay di erential-di erence systems with time-varying delays. As a concluding remarkwe also show how, in discrete-time, some peculiar connections withswitched systems in block-companion form can be derived, suggest-ing future interplay with recent results on identi cation of switchingsystems via machine learning techniques.

“Sampled-Data Control of Nonlinear Time-Delay Systems”

It is well known that, nowadays, the digital implementation of model-basedcontrol laws is universally used in practice. A crucial ingredient in practical control of a(technological, physical, chemical, biological) system is an understanding of the impact ofsampling on continuous-time models describing the reality in exam. As well, retarded functionaldi erential equations describe a signi cant quantity of practical nonlinear systems (see, amongthe others, chemical reactors, glucose-insulin models in humans, tele-manipulated robots, high-speed Internet congestion, neural networks, biological systems and population dynamics).Nevertheless, few results are available in the literature concerning sampled-data control ofthe above systems, given the inherent diculties coming from their concurrent nonlinear andin nite dimensional character. In this seminar, new results concerning the sampled-data controlof a general class of nonlinear systems with state-delays (i.e., systems described by retardedfunctional di erential equations) are shown. In particular, the stability preservation problemunder emulation and quantization of continuous-time controllers for nonlinear time-delaysystems is addressed. Practical and exponential stability preservation results under samplingare provided. In the presented results, possible actuation disturbances and observation errorsare considered. Tools for the design of robust sampled-data controllers are provided. Inputsaturation constraints are also managed. The stabilization in the sample-and-hold sense theory,the converse Lyapunov theorems and the Lyapunov redesign methods are used as tools toprove the results. Applications, concerning the regulation problem of plasma glycemia in type2 diabetic patients, are also presented.

“Efficient data transfer and networking for the Internet of Things”

In this seminar, we review novel techniques and recent advancements in short range radio standard solutions for Internet of Things applications, focusing on the Bluetooth technology and the IEEE 802 family of standards. In the first part, we assess the performance of multiple data transfer modes specified in the latest version of Bluetooth, comparing the two asynchronous advertising modes and the baseline connection-oriented mode. Service ratio, communication delays, and battery lifetime are evaluated in a small-scale home automation use case with a realistic deployment and varying traffic loads. In the second part, a networking solution based on the recently released Bluetooth Mesh Profile specification is illustrated and its performance evaluated in a large-scale scenario. Results show that with proper deployment and configuration of relevant parameters of the protocol stack, the proposed mesh networking solution is able to support the operation of dense networks with thousands of devices, providing the necessary robustness and service ratio. Furthermore, the network design is flexible enough to handle the introduction of managed operations on top of a flooding-based forwarding mechanism, to further optimize networking procedures and automate the relay selection process.

“Decision and control tools for sustainable management of logistics and transportation systems”

Transport is one of the most energy consuming and polluting activities and its environmental impact is expected to furtherly grow in the future. Nonetheless it is an indispensable activity for the economic and social development of industrialized countries. As a consequence there is a growing need for decision and control tools to allow managing and optimizing in a more efficient and sustainable manner the transport sector.

Starting from the above, and taking into account the suggestions provided by the European Commission and by the research community, this seminar proposes some possible actions to optimize the transport sector in order to reduce its environmental impact. In particular, three main sub-themes are focused: the logistics (both in terms of internal logistics and warehouses analysis and optimization, and external logistics and Supply Chain Networks design), the freight transportation (particularly focusing on multimodal transport of goods and intermodal rail-road terminals operations), and the passenger transportation (with some proposals to reduce the traffic congestion in metropolitan and urban area). Finally, since in the modern smart city context it is also crucial to ensure a sustainable governance, a multi-criteria decision making approach to select the optimal strategy for the smart development of a metropolitan area in a set of candidate action plans under uncertain data is also proposed.

“Nowcasting Monthly GDP with Big Data: a Model Averaging Approach”

Gross domestic product (GDP) is the most comprehensive and authoritative measure of economicactivity. The macroeconomic literature has focused on nowcasting and forecasting this measureat the monthly frequency, using related high frequency indicators. The paper addresses the issueof estimating monthly gross domestic product using a large dimensional set of monthly indicators,by pooling the disaggregate estimates arising from simple and feasible bivariate models thatconsider one indicator at a time, in conjunction to GDP or a component of GDP. The weightsused for the combination reect the ability to nowcast the original quarterly GDP component.Our base model handles mixed frequency data and ragged-edge data structure with any patternof missingness. Our methodology allows to assess the contribution of the monthly indicators tothe estimation of monthly GDP, thereby providing essential information on their relevance. Thisevaluation leads to several interesting discoveries.

“Parallel and distributed convergent decomposition algorithms for the training of SVMs”

Support Vector Machines (SVMs) are well known supervised learning techniques widely used for linear and nonlinear binary classification. The training of SVMs can be formulated as a convex constrained optimiza- tion problem, where a quadratic objective function is minimized over a feasible set made up of linear constraints. The dimension of the SVM training problem is equivalent to the number of samples of the considered training set. Hence, for many instances of practical interest the training problem cannot be efficiently solved by means of standard optimization techniques. To cope with these instances a decompositon approach is commonly adopted. Decomposition methods split the original problem into a sequence of smaller subproblems in which only a subset of  vari- ables is optimized at a time. For big and huge dimensional training sets decomposition should be combined with the simultaneous parallel or dis- tributed optimization of multiple subproblems. However, the design of practically and theoretically convergent parallel and distributed decom- position algorithms is a complicated task. Indeed, the parallel SVMs training algorithms reported in the literature, either are heuristics con- verging to a sub-optimal solution, or are not provided with convergence properties. In this seminar a parallel and a distributed convergent decom- position algorithmic schemes are presented. Their convergence properties can be proved under very mild assumptions and exploiting simple ideas.

Related topics