Este projeto consiste em aplicar o Reinforcement Learning a problemas de otimização, mais concretamente problemas de gestão de stock, sendo que o objetivo principal é determinar as quantidades de stock a encomendar ideais, de forma a minimizar os custos que o armazenamento e a rutura de stock acarretam. Este método é comparado com métodos mais tradicionais de gestão de stock, de forma a identificar as vantagens e desvantagens de cada abordagem.

The objective of this report is to implement a Convolutional Neural Network (CNN) in an FPGA, with a main focus on accelerating the training, using Maxeler technology as a way to compile higher level code directly into hardware.Neural Networks are one of the most commonly used models used in all sorts of tasks in Machine Learning. This type of network is mostly used for image recognition/generation, since a few layers ( convolutional, pooling) can be viewed as image operations to find features, which are then combined in the fully connected layer(s) and used to produce the output.

No description available.

No description available.

A utilização acentuada de dispositivos móveis, as exigências de uma boa Qualidade de Serviço (Quality of Service - QoS) aos utilizadores e o aumento do número de Eventos Temporários com Muita Gente (Temporary Crowded Events - TCEs) nos dias de hoje remetem para a utilização de redes aéreas para servir estas necessidades e novas tendências.Uma Rede Aérea Multicamada com Reconhecimento de Tráfego (Traffic-Aware Multi-tier Flying Network - TMFN) é uma arquitectura de dois níveis, onde são utilizados Ponto de Acesso de Rede Voadora (Flying Mesh Access Point - FMAP) que funcionam como pontos de acesso móveis para uma adaptação do serviço à disposição dos utilizadores no terreno destes eventos. O algoritmo de Network Planning (NetPlan) é aplicado em cima desta arquitectura e pressupõe uma reconfiguração da topologia Veículo Aéreo Não Tripulado (Unmanned Aerial Vehicle - UAV) que garanta um bom QoS aos utilizadores.Porém, esta dissertação procura definir e construir o modelo de Aprendizagem de Máquina (Machine Learning - ML) que, aplicado a esta arquitectura de rede, permita obter resultados melhores que aqueles obtidos pelo NetPlan, ao procurar as soluções ótimas para todo o tipo de cenários, incluindo cenários dinâmicos. Para tal, será usado uma abordagem de Aprendizagem Reforçada Profundamente (Deep Reinforcement Learning - DLR). Esta abordagem passa pela interseção de um algoritmo de Aprendizagem Reforçada (Reinforcement Lerning - RL), que através da definição de uma função objetivo que, é recompensada ou penalizada consoante o resultado de determinadas ações no ambiente, irá aprender de maneira iterativa a adaptar-se para os diferentes cenários, com técnicas de Aprendizagem Profunda (Deep Learning - DL) que extraem características intrínsecas à rede/ambiente, de modo a melhorar a construção da função objetivo. / Nowadays, the strong use of mobile devices, the demands of a good Quality of Service (QoS) given to the users and the increase in the number of Temporary Crowded Events (TCEs) are leading to the use of flying networks to serve these needs and new trends.A Traffic-Aware Multi-Tier Flying Network (TMFN) is a two-level architecture, where Flying Mesh Access Points (FMAPs) are used as mobile access points for adapting the service to the users' disposal in the field of these events. The Network Planning algorithm is applied on top of this architecture and presupposes a reconfiguration of the Unmanned Aerial Vehicle (UAV) topology that guarantees a good QoS to the users.This dissertation seeks to define and build an ML model that, applied to this network architecture, allows better results than those obtained by NetPlan, when searching for optimal solutions for all types of scenarios, including dynamic scenarios. To do so, a Deep Reinforcement Learning (DRL) approach will be used. This approach passes through the intersection of an Reinforcement Learning (RL) algorithm, which, through the definition of an objective function that is rewarded or penalised according to the result of certain actions in the environment, will iteratively learn the best way to adapt itself to different scenarios, with Deep Learning (DL) techniques, that extracts inherent characteristics of the network / environment, in order to improve the development of the objective reward function.

No description available.

No description available.

The sport has been a part of our lives since the beginning of times, whether we are spectators or participants. The diffusion and increase of multimedia platforms made the consumption of these contents available to everyone. Sports videos appeal to a large population all around the world and have become an important form of multimedia content that is streamed over the Internet and television networks. Moreover, sport content creators want to provide the users with relevant information such as live commentary, summarization of the games in form of text or video using automatic tools.As a result, MOG-Technologies wants to create a tool capable of summarizing football matches based on semantic content, and this problem was explored in the scope of this Dissertation. The main objective is to convert the television football commentator's speech into text taking advantage of Google's Speech-to-Text tool. Several machine learning models were then tested to classify sentences into important events. For the model training, a dataset was created, combining 43 games transcription from different television channels also from 72 games provided by Google Search timeline commentary, the combined dataset contains 3260 sentences. To validate the proposed solution the accuracy and f1 score were extracted for each machine learning model.The results show that the developed tool is capable of predicting events in live events, with low error rate. Also, combining multiple sources, not only the sport commentator speech, will help to increase the performance of the tool. It is important to notice that the dataset created during this Dissertation will allow MOG-Technologies to expand and perfect the concept discussed in this project.

The main objective of this project is the elaboration and testing of a Power Transformer Overload Capacity model/algorithm. Starting from a desired load factor or overload time, based on the initial thermal conditions, the algorithm returns the maximum overload time or load factor that respects the thermal limits of the transformer.Taking into consideration the practices of the institution, with the actual data provided by all the engineering work, some suggestions are made regarding the limitations found in relation to the intended algorithm.

É cada vez mais evidente a importância e a necessidade, em empresas de topo relacionadas com gestão, de sistemas de apoio à decisão que auxiliem em problemas de roteamento e de setorização.Nesse sentido, nesta dissertação, são abordadas várias soluções para resolução desses problemas, no que diz respeito ao caso de estudo em concreto sobre recolha de pacientes, bem como o seu transporte das suas residências para o centro hospitalar em questão, onde o alvo é a região do Grande Porto, em que se procura obter a menor distância percorrida possível (sendo possível, também, analisar em termos de otimização de custos ou recursos), obedecendo a um conjunto de restrições, restrições essas que podem passar por caraterísticas físicas ou psicológicas específicas de cada paciente, localização geográfica dos utentes em questão, entre outras.Daí, nascem as heurísticas e meta-heurísticas, de forma a poder incluir qualquer tipo de condições que sejam necessárias para o problema em questão.Assim, é pretendido setorizar a área do Grande Porto, para um conjunto de pacientes, mediante a sua localização, bem como definir a melhor rota a ser percorrida, tendo em conta a capacidade do veículo de recolha e transporte e eventuais restrições que dizem respeito ao paciente em si. / The importance and necessity of top management-related decision support systems that assist in routing and sectorization problems is increasingly evident.In this sense, in this dissertation, several solutions are addressed to solve these problems, with regard to the concrete case study on patient collection, as well as their transportation from their residences to the hospital center in question, where the target is the region of Porto, where it is ideal to obtain the shortest possible distance (it is also possible to analyze in terms of minimization of costs or resources), obeying a set of restrictions, which may be specific physical or psychological characteristics of each patient, geographical location of the users in question, among others.Hence, heuristics and metaheuristics are born, so that it can include any kind of conditions that are necessary for the problem in question.Thus, it is intended to segment the Porto area, for a set of patients, by location, as well as defining the best route to be traveled, taking into account the vehicle's collection and transportation capacity and possible restrictions which relate to the patient itself.