409 research outputs found
Sort by
Automatic and Accurate Performance Prediction in Distributed Systems
Full text linkSystem performance is getting attention by industry as it affects
user experience, and much research focused on performance
evaluation approaches. Profiling is the most straightforward
approach to performance evaluation of software systems,
despite being limited to shallow analyses. Conversely,
software performance models excel in representing complex
interactions between components. Still, practitioners do not
integrate performance models in the software development
cycle, as the learning curve is too steep, and the approaches
do not adapt well to incremental development practices. In
this thesis, we propose three approaches towards automatic
learning of performance models. The first approach employs
a Recurrent Neural Network (RNN) to extract a full Queueing
Network (QN) model of the system; the second one calibrates
a Layered Queueing Network (LQN) using an RNN;
the third one presents μP, a framework that allows the user
to develop microservice systems and obtain the corresponding
LQN model from source code analysis. We considered
the microservices architecture as it is embraced by influential
players (e.g., Amazon, Netflix). Those approaches have
two advantages: i) minimal user intervention to flatten the
learning curve; ii) continuous synchronization between software
and performance model, such as each software development
iteration is reflected on the model. We validated our
approaches on several benchmarks taken from the literature.
The models we generate can be queried to predict the system
behavior under conditions significantly different from
the learning setting, and the results show sensible advancements
in the quality of the predictions
Learning-based Stochastic Model Predictive Control for Autonomous Driving
Full text linkAutonomous driving in urban environments requires safe control
policies that account for the non-determinism of moving
obstacles, for instance, the intention of other vehicles while
crossing an uncontrolled intersection. This thesis addresses
the aforementioned problem by proposing a stochastic model
predictive control (SMPC) approach. In this approach, we
consider robust collision avoidance as a constraint to guarantee
safety and a stochastic performance index that will increase
the quality of the closed-loop tracking by ignoring the
unlikely obstacle configurations that could occur. We compute
the probabilities associated with different obstacle trajectories
by training a classifier on a realistic dataset generated
by the microscopic traffic simulator SUMO and show the
benefits of the proposed stochastic MPC formulation in a simulated
real intersection. This thesis is divided into two parts:
first, discuss the formulation of the existing control algorithm
and our proposed approach, and second, the scenario prediction
of the obstacle vehicles
At the intersection between machine learning and econometrics: theory and applications
Full text linkIn the present work, we introduce theoretical and application novelties at
the intersection between machine learning and econometrics in social and
health sciences. In particular, Part 1 delves into optimizing the data collection
process in a specific statistical model, commonly used in econometrics,
employing an optimization criterion inspired by machine learning, namely,
the generalization error conditioned on the training input data. In the first
Chapter, we analyze and optimize the trade-off between sample size, the precision
of supervision on a variation of the unbalanced fixed effects panel data
model. In the second Chapter we extend the analysis to the Fixed Effects
GLS (FEGLS) case in order to account for the heterogeneity in the data
associated with different units, for which correlated measurement errors corrupt
distinct observations related to the same unit. In Part 2, we introduce
applications of innovative econometrics and machine learning techniques. In
the third Chapter we propose a novel methodology to explore the effect of
market size on market innovation in the Pharmaceutical industry. Finally, in
the fourth Chapter, we innovate the literature on the economic complexity
of countries through machine learning. The Dissertation contributes to the
literature on machine learning and applied econometrics mainly by: (i) extending
the current framework to novel scenarios and applications (Chapter
1 - Chapter 2); (ii) developing a novel econometric methodology to assess
long-debated issues in literature (Chapter 3); (iii) constructing a novel index
of economic complexity through machine learning (Chapter 4)
Essays on Innovation Networks and Global Cities
Full text linkIn a competitive economy, technological innovation is a core element of economic growth and development, and its accumulation in a rapidly changing technological environment is key to adaptation. This thesis investigates how cities, particularly global cities can develop new technological capabilities to enter new technological fields and become competitive in new technological areas. First, a new geo-referenced patent database is developed to overcome some limitations of the existing international patent repository to make an international comparison of cities feasible. The new database provides broad coverage of cities in developed and emerging economies and allows us to tackle the following research questions: (i)how the co-patenting network of domestic and international linkages of cities has changed over time? (ii) which cities are more technologically complex? Can we predict the future evolution of cities’ technological complexity? (iii) What is the effect of inter-city linkages and technological relatedness
on the likelihood of cities to enter new technological areas? (iv) which cities are the most central in coordinating complex international teams of inventors on a global scale?
To geo-locate patents, an address retrieval algorithm has been applied to resolve the problem of missing addresses, exploiting the availability of address in patent family and similarity of inventors based on attributes such as working for the same applicant. A harmonized definition for functional urban areas is used to assign patents to cities on a global scale. The database provides a significant improvement from the raw PATSTAT dataset with an estimated confidence level of 84-88%.
We analyze both unweighted (extensive) and patent-weighted (intensive) linkages between cities to address the first research question. The preliminary findings show an increase of international ties both intensively and extensively and the reliance of cities in developing economies on international ties to global cities. To address the second research question, we apply the generalized economic complexity (GENEPY) algorithm to measure the economic
complexity of cities based on patent production in these cities. We use machine learning models (Random Forest, XGBoost, SVM, Neural network) to forecast the future technological complexities of cities. We show that the machine learning models (especially Random Forests) have higher predictive power than the benchmark model (time-independent conditional probabilities) as they account for higher-order and non-linear interdependencies between technologies.
To address the third question, we applied a stratified semiparametric Cox proportional hazard model to examine the
likelihood of cities entering new technologies. We show that inter-city linkages and technological relatedness significantly increase the likelihood of entry into new technological areas.
Inter-city linkages are more critical for non-global cities than for global cities to enter new technological areas, whereas linkages to inventors located in cities with a large pool of inventors positively moderate the effect of inter-city linkages on entry. For the last research question, we use hypergraphs structure and propose a measure based on 3-hyperedges (three cities in multiple countries) in the collaboration networks constructed
from scientific publications and patents to identify the most competitive global cities in the international network of inventors. To this end, we construct a null model using the hypergeometric ensembles of random graphs and find that five US cities play a leading role in transnational networks of researchers. San Francisco stands out as the most global city, but Shanghai is rapidly emerging as a global player
Essays in Applied Economics
No full textThe present thesis consists of three independent chapters. The frst chapter analyses the effect of religiosity on innovativeness. The empirical literature on the relation between religion and innovation hitherto applied regressions without considering endogeneity in the estimates, raising questions about spurious correlations. This chapter provides the frst empirical study to build a causal link between religion and innovation by employing the instrumental variables method to untangle possible endogeneity. The results strongly suggest that higher religiosity has a somewhat negative effect on innovativeness. Three possible causality channels from religiosity to innovation are discussed: time allocation argument, the fear of uncertainty, and traditional roles empowered by religion. The second chapter explores the importance of regional
capabilities, in the form of workplace skills, in the industrial diversifcation process of regions by exploiting two recently developed approaches: relatedness and economic complexity. Building on the network-based approach of evolutionary economic geography, the study shows that
workplace skills form two highly polarised clusters into social-cognitive and technical-physical skills. The econometric analysis indicates that industries have a higher (lower) probability of developing a comparative advantage if their required skill set is (not) similar to those available
in the region, regardless of the skill type. Nevertheless, similarity to technical-physical skills and higher complexity in social cognitive skills yield the highest regional competitive advantage probabilities. The third chapter analyses the Ramsey pricing of pharmaceuticals by using recently developed the debiased/double orthogonal machine learning method that allows for heterogeneous treatments in a dynamic panel setting. The study assesses the validity of the inverse elasticity rule by providing a cross-country analysis of pharmaceutical demand at the
molecule level. The results show that pharmaceutical prices vary inversely with price elasticities, both in high-income and low-middle-income countries, signalling the existence of Ramsey pricing
Religious Cultural Heritage and Political Contestation: the Role of UNESCO The Cases of the Old City of Hebron, the Four Medieval Monuments in Kosovo and Metohija and the Temple of Preah Vihear
Full text linkReligious cultural heritage listed by UNESCO as World Heritage Sites or World Heritage Sites in Danger is chosen as a leading category of cultural heritage, taking into consideration that it is the object of claims
between States. These two aspects are the ground for political claims on contested territories. Specific cases of contested religious heritage rooted in contested territories, such as the Palestinian-Israeli, Kosovo-Serbian
and Cambodian-Thailand cases, are the object of the current
examination.
In particular, three elements will be taken into consideration: a) how UNESCO and National Governments build the Outstanding Universal Value of these selected cases; b) how UNESCO considers the element of contestation in the decisional-making process related to the listing
processes of contested sites; c) how intangible heritage has a relevance as a tool to enforce the political claims of parties. In this perspective, UNESCO’s decisions seem to have different impacts (enforcement of national states/ political entities, impairment of national states/political entities, definition of borders1) while managing conflicts of sovereignties between States through religious cultural heritage
Learning optimal control policies from data: a partially model-based actor-only approach
Full text linkThis dissertation presents new algorithms for learning optimal feed-back controllers directly from experimental data, considering the plant to be controlled as a black-box source of streaming input and output data. The presented methods fall in the Reinforcement Learning “actor-only” family of algorithms, employing a represen-tation (policy parameterization) of the controller as a function of the feedback values and of a set of parameters to be tuned. The
optimization of a policy parameterization corresponds to the search of the set of parameters associated with the best value of a chosen performance index. Such a search is carried on via numerical opti-mization techniques, such as the Stochastic Gradient Descent algo-rithm and related techniques. The proposed methods are based on a combination of the data-driven policy search framework with some
elements of the model-based scenario, in order to mitigate some of the drawbacks presented by the purely data-driven approach, while retaining a low modeling effort, as compared to the typical identif-cation and model-based control design scenario.
In particular, we initially introduce an algorithm for the search of smooth control policies, considering both the online scenario (when new data are collected from the plant during the iterative policy syn-thesis, while the plant is also under closed-loop control) and the of-fine one (i.e. from open-loop data that were previously collected from the plant). The proposed method is then extended to learn non-smooth control policies, in particular hybrid control laws, op-timizing both the local controllers and the switching law directly from data. The described methods are then extended in order to be employed in a collaborative learning setup, considering multi-agent systems characterized by heavy similarities, exploiting a cloud-aided scenario to enhance the learning process by sharing information
Essays on firms' competitiveness
Full text linkAs competition becomes intense, firms seek strategies to keep afloat. Among others, they carefully choose their activity location, recruit a talented workforce, and engage in innovation. In this thesis, we shed light on these three features empirically, mainly using econometric techniques. Our contribution is in the literature of firms’ competitiveness, industrial organization and economic geography. At first, we study regional productivity disparities and their interplay with local agglomeration advantages. To do so, we apply a density-based machine learning clustering algorithm to identify firms’ clusters at a fine-grained geographic scale on a sample of Italian firms. Then, we observe simultaneously the extent to which clusters explain agglomeration economies and firm selection effects. Our findings suggest that dense clusters generate agglomeration externalities that are heterogeneous across regions. In the second part of the thesis, we
investigate the impact of foreign managers on firms’ competitiveness on a sample of firms operating in the United Kingdom. We show that domestic firms become more efficient
after recruiting foreigners to their management team due to
previous industry-specific experience. In the last part, we assess the impact of patents on market share and labour productivity in the global Information and Communication Technologies (ICT) sector. Using a recent difference-in-difference approach, we find that patenting increase market share without significantly affecting labour productivity. Our evidence indicates some concerns regarding the implications of property rights from innovation on market competition
Optimizing complex networks models
Full text linkAnalyzing real-world networks ultimately amounts at com-
paring their empirical properties with the outcome of a proper, statistical model. The far most common, and most useful, approach to define benchmarks rests upon the so-called canonical formalism of statistical mechanics which has led to the definition of the broad class of models known as Exponential Random Graphs (ERGs). Generally speaking, employing a model of this family boils down at maximizing a likelihood function that embodies the available information about a certain system, hence constituting the desired benchmark. Although powerful, the aforementioned models cannot be solved analytically, whence the need to rest upon numerical recipes for their optimization. Generally speaking, this is a hard task, since real-world networks can be enormous in size (for example, consisting of billions of nodes and links), hence requiring models with ‘many’ parameters (say, of the same order of magnitude of the number of nodes). This evidence calls for optimization algorithms which are both fast and scalable: the collection of works constituting the present thesis represents an attempt to fill this gap. Chapter 1 provides a
quick introduction to the topic. Chapter 2 deals specifically with ERGs: after reviewing the basic concepts constituting the pillars upon which such a framework is based, we will discuss several instances of it and three different numerical techniques for their optimization. Chapter 3, instead, focuses on the detection of mesoscale structures and, in particular, on the formalism based upon surprise: as the latter allows any partition of nodes to be assigned a p-value, detecting a specific, mesoscale structural organization can be understood as the problem of finding the corresponding, most significant partition - i.e. an optimization problem whose score function is, precisely, surprise. Finally, chapter 4 deals with the application of a couple of ERGs and of the surprise-based formalism to cryptocurrencies (specifically, Bitcoin)
Network analysis of a complex disease: the gut microbiota in the inflammatory bowel disease case
Full text linkThe gut microbiota contains hundreds of types of microbes
and dysbiosis can lead to infammatory bowel diseases (IBD),
which comprise Crohn’s disease (CD) and ulcerative colitis
(UC). Due to the complex nature of the IBD, it is interest-
ing to understand the differences between a control (NI) and
an IBD gut microbiome by using new tools offered by net-
work science. In particular, when metagenomic data are con-
sidered, it is possible to build networks according to the co-variance, the co-occurrence and multiple layers of networks(multilayer networks). In addition to the construction of the
networks, an analysis of the differential expressed pathways
is carried out, several centrality measures are calculated, and community detection is performed to explore the topological differences between the diagnosis networks. The analysis of the correlation network topology highlights that, in IBD net-works, the pathway involving coenzyme A of the unclassifed species becomes central. Furthermore, the modularity in the IBD networks is higher. In both the correlation network and the co-occurrence network, the modules belonging to B. ova-tus and B. caccae are positioned differently in each diagnosis.
Furthermore, the difference between the NI and the UC diag-
nosis networks lies in a change in the wiring that preserves
the centralities. Moreover, the fundamental role of two of the Roseburia species in the NI is evidenced. A further step will consist of identifying the minimum number of pathways on which it would be ideal to intervene to drive the system back to a healthy state by the precision medicine way of operating