Ou procure na linha do tempo com mais detalhes.

Automated Planning to Support the Deployment and Management of Applications in Cloud Environments

Cloud computing has enabled myriad of applications to benefit from dynamic provisioning of resources. These applications must adjust their resources in response to changes in the environment in order to satisfy business-defined goals about service quality. However, engineering decision-making mechanisms to help the deployment and management of resources in cloud contexts is a challenging task. In fact, cloud applications call for automated mechanisms that: (1) explore efficiently large solution spaces (defined by the combination of machine types, provisioning actions, and state transitions expected in the temporal horizon); (2) generate deliberate plans to operate the system in a way that satisfies requirements, maximizes performance and minimizes operational costs; and (3) support the definition and revision of policies to adapt the system under expected conditions. Automated Planning, the area of artificial intelligence concerned with synthesizing plans of actions to achieve a goal, offers opportunities to address these challenges. This thesis focuses on the design and evaluation of approaches that exploit automated planning to support the deployment and management of applications running in cloud environments. To this purpose, this thesis presents three contributions: (1) a solution to the (offline) generation of reactive policies, that exploits temporal planning languages and tools to support the definition and revision of policies, applicable under common conditions; (2) a solution to the (online) generation of proactive plans, that takes advantage of longterm temporal planning and behavioral predictions to reconfigure interactive applications; and (3) a solution to the (offline) generation of execution policies, that resorts to probabilistic planning to deal with the uncertainty caused by spot instance revocations in the deployment of workflow applications. These proposals have been evaluated using realistic case studies of elastic scaling and workflow executions in the cloud. Results support the claim that automated decision-making mechanisms that rely on planning are scalable and responsive, and able to guide the system to satisfy requirements, optimize performance and minimize operational costs.

access_time 28 de fevereiro de 2019 às 16:00
place Room 0.19, Pav. Inf. II, IST, Alameda
local_offer Prova de Doutoramento
person Candidato: Richard Joaquin Gil Martinez N.º 79788
supervisor_account Orientador: Prof. Luís Eduardo Teixeira Rodrigues

Security vs. Maintainability: Fixing Vulnerabilities Obfuscates your Code

Security is a crucial non-functionality requirement for software applications. However, building secure software is far from trivial as developers lack both the knowledge and tools to effectively address this concern. In this paper, we study the impact of changes to improve security on the maintainability of several open source applications. Using a dataset containing 607 security- oriented commits, we measure maintainability — as computed by the Software Improvement Group’s web-based source code analysis service Better Code Hub (BCH) — before and after the security refactoring. Results show that making software more secure comes at a cost on maintainability. This is particularly evident in refactorings to deal with Broken Authentication and Cross-Site Request Forgery attacks.

access_time 07 de fevereiro de 2019 às 14:00
place INESC-ID Room 418
local_offer Tópicos de Investigação
person Candidato: Sofia Oliveira Reis N.º 91520
supervisor_account Orientador: Prof. Rui Filipe Lima Maranhão de Abreu

Cross-lingual prosody evaluation through transfer learning

The importance of automatic prosody assessment has been acknowledged, as it provides relevant information about the speaker, the languages, the pragmatics and paralinguistics of speech. Nevertheless, available data sets are often insufficient for the tasks aimed at, namely when it comes to the usage of DNNs, which require great amounts of data. Transfer learning is a state-of-the-art technique being used for several tasks and proven to be very informative, as training with diverse data sets and testing on distinct ones can assure robustness and cross-lingual analysis. In this case, we investigate whether transfer learning can be applied to L2 learners, either exclusively in learning contexts or in e-health ones too, which is a very challenging task.The main data set we use was built for an intonation imitation task with native speakers of Portuguese, whose assessment relied on a DTW algorithm only. Building upon previous work, a nativeness assessment task of L2 speakers of English, we applied a DNN for feature extraction. We have considered features with no temporal dependency and features with temporal dependency, corresponding to an LSTM layer.

access_time 07 de fevereiro de 2019 às 11:00
place INESC-ID, Room 336
local_offer Tópicos de Investigação
person Candidato: Mariana Dimas Julião Nº 94140
supervisor_account Orientador: Prof. Alberto Abad Gareta / Dra. Helena Moniz

Group-based Adaptation for Learning: Defining how serious games adapt over specific learning moments

Due to the influence of education in the development of society, several methods have been proposed to improve learning. In particular, digital Adaptive Learning Systems (ALSs) have been developed to tailor content to the students’ needs. In this work, we define Group-based Adaptation for Learning (GAL), a methodology which aims to automatically adjust game mechanics based on the characteristics of a group of students so that their collective ability is improved. A complete validation of our model would require its application to a suitable game and the execution of user studies. In this paper, we present several agent-based simulations that attempt to validate certain aspects of the model and inform us about additional refinements and improvements before exploring such route in the future. More specifically, we compared the impact of our group formation strategy as opposed to an optimal and a baseline random one. We checked that the average ability increases and engagement of the students rapidly converged to high, near optimal values when using GAL, as opposed to the random strategy which maintained low values. Moreover, unlike the random strategy, GAL managed to considerably approximate the promoted learning profiles to the preferences of the students. Finally, we describe how our method could function in a specific domain by defining a music serious game which allows students to learn to play instruments.

access_time 07 de fevereiro de 2019 às 11:00
place IST Taguspark, Room 2N7.1
local_offer Tópicos de Investigação
person Candidato: Samuel Gomes
supervisor_account Orientador 1: Prof. João Dias
supervisor_account Orientador 2: Prof. Carlos Martinho

The Impact of Virtual Reality in the Social Presence of a Virtual Agent

In this work we hypothesize that having a virtual character in Virtual Reality will have a stronger impact on the human interactor compared to traditional 3D computer application. In order to test our theory we conducted an experiment with the same human-virtual agent interaction between a digital environment (in front of a computer screen) and a virtual reality one. The scenario used was a police interrogation where the player performed the Interrogator role whose objective was to get as much information as he could out of a Virtual Agent Suspect. Results across 31 different test subjects confirmed that users felt that the Virtual Agent they were interacting with had a higher Social Presence in Virtual Reality when compared to a computer screen.

access_time 07 de fevereiro de 2019 às 10:00
place IST Taguspark, Room 2N7.1
local_offer Tópicos de Investigação
person Candidato: Manuel Guimarães
supervisor_account Orientador 1: Prof. João Dias
supervisor_account Orientador 2: Prof. Pedro Santos

Navegue pelo calendário para descobrir os eventos.