ISUAM: Intelligent and Safe UAM with Deep Reinforcement Learning

/Robot_ensemble: A Social Layer for Service Robots Using Component Ensembles Paradigm

/4DNavMCTS: a novel approach of conflict detection and resolution for 4-dimensional trajectory-based operation

/Influence analysis between logical and physical models in document-oriented NoSQL databases

/Deep-Quality-EON Classifier and DRL approach

/LayoutQT - Layout Quadrant Tags to embed Visual Features for Document Analysis

 

Horário: 14h

Palestrante: Cristiano Perez Garcia (mestrado) 

Orientador: Prof Li Weigang

Title: ISUAM: Intelligent and Safe UAM with Deep Reinforcement Learning

Resumo: O espaço aéreo é um recurso limitado e seu uso tende a crescer de forma substancial com o desenvolvimento as aeronaves e-VTOL. O grande número de veículos concentrados em regiões urbanas traz a necessidade de atenção em relação a potenciais conflitos. Tal tarefa exige grande capacidade computacional e o emprego de modelos cada vez mais sofisticados. Os modelos de aprendizado profundo são capazes de reconhecer padrões em situações complexas e podem auxiliar a encontrar formas mais eficientes de resolução de conflitos quando o objetivo é minimizar o tempo adicional gasto nos desvios para evitar um conflito. O objetivo deste trabalho é o desenvolvimento de um modelo de aprendizado de reforço profundo capaz de encontrar soluções para conflitos entre aeronaves no ambiente UAM de forma eficiente e segura.

 

Horário: 14h20

Palestrante: Gabriel Siqueira Rodrigues (doutorado) 

Orientadora: Profa Genaína Rodrigues

Title: Robot_ensemble: A Social Layer for Service Robots Using Component Ensembles Paradigm

Abstract: Service robots applications are supposed to operate in an indoor environment, such as a hospital or restaurant, with multiple robots executing task oriented missions, while sharing the environment with humans, other robots and connected devices. In such environment, a number of collaborations can be enabled by communications, and can take different form such as requests of service, sharing of knowledge, and resources negotiation. Organizing such communication can be challenging as the environment is dynamic (agents can enter and live) and mobile robots can experience network outage while moving through blind stops in the environment. In such a context a non-principled design, with network call scattered across the robot control logic can lead to failures that are difficult to fix. In this paper we propose a social layer using the component ensembles paradigm that abstracts the communication in a MRS, allowing robots to collaborate with users and devices in the environment based on ensembles. Firstly, we review the main concepts of ensemble based systems, previous works in ECS and limitation for its application in the service robots domain. Secondly, we present our analysis of how ensemble can generalize collaborations in a service robots environment, and our implementation on top of the ROS middleware. Finally, we present an experiment in a simulated environments.

 

Horário: 14h40

Palestrante: Lucas Borges Monteiro (doutorado)

Orientador: Prof Li Weigang

Title: 4DNavMCTS: a novel approach of conflict detection and resolution for 4-dimensional trajectory-based operation

Abstract: Technological progress has greatly increased the amount of data generated by various fields, including air transportation. Correctly handling this massive data can bring important results because it makes decision-making more accurate. In this sense, with a focus on Trajectory Based Operations (TBO) in Air Traffic Management (ATM), a modeling system based on Monte Carlo Tree Search (MCTS) is developed. Named 4DNavMCTS, this novel model of four-dimensional trajectory representation can satisfactorily deal with complex TBO problems. Considering the navigation big data, 4DNavMCTS can perform conflict detection and resolution (CD&R) under the artificial intelligence (AI) paradigm to find and solve the potential conflicts between aircraft and improve flight safety with reasonable prediction. Furthermore, using the vector space model (VSM) to represent the possible strategies, the proposed approach can help controllers choose the appropriate trajectory quantitatively and reduce the risk of conflict. The experiment results from the case study demonstrated the effectiveness of the newly developed model in the real scenario of CD&R.

 

Horário: 15h

Palestrante: Guilherme Enéas Vaz Silva (doutorado) 

Orientador: Prof  André Drummond

Title: Deep-Quality-EON Classifier and DRL approach

Abstract: Deep-Quality-EON Classifier and DRL approach

 

Horário: 15h20

Palestrante: Harley Vera Olivera (doutorado) 

Orientadora: Profa Maristela Holanda

Title: Influence analysis between logical and physical models in document-oriented NoSQL databases

Abstract: This work analyzes the relationship between logical and physical models in document-oriented NoSQL data models. In particular, we tend to show that the performance of a logical model will be the same regardless of the physical model used. For this purpose, three logical models (referenced, embedded, and hybrid) are analyzed into three physical models (standalone, replica set, and sharding). Results show that the performance of the logical model tends to be the same in any physical model.

 

Horário: 15h40

Palestrante: Patricia Medyna Lauritzen de L. Drumond (doutorado) 

Orientador: Prof Teófilo Campos

Title: LayoutQT - Layout Quadrant Tags to embed Visual Features for Document Analysis

Abstract: The relative position of text blocks plays a crucial role in document understanding. However, the task of embedding layout information in the representation of a page instance is not trivial. Computer Vision and Natural Language Processing techniques have been advancing on the extraction of content from document images considering layout features. We propose a set of Layout Quadrant Tags (LayoutQT) as a new way of encoding layout information in textual embedding. We show that this enables that a standard NLP pipeline be significantly enhanced without requiring expensive midor high-level multimodal fusion. We evaluate our method on two datasets, Tobacco-800 and RVL-CDIP, for document image classification tasks. The document classification performed with our method obtained an accuracy of 83.6% on the large-scale RVL-CDIP and 99.5% on the Tobacco 800 datasets.

 

Local: Teams MS - Equipe PPGI-316415 Seminário, Canal Seminários 1-2022

 

https://teams.microsoft.com/l/channel/19%3a70ff576247034ebba160e70b2d691d93%40thread.tacv2/Semin%25C3%25A1rios%25201-2022?groupId=93b66213-b249-467a-bcbe-dcd4255edf95&tenantId=ec359ba1-630b-4d2b-b833-c8e6d48f8059

 

Profa Célia Ghedini Ralha (Este endereço de email está sendo protegido de spambots. Você precisa do JavaScript ativado para vê-lo.)

Coordenadora Seminários de Pós-Graduação em Informática 1-2022