23-04-2001
Inzendingen VVS-scriptieprijs voor Statistiek en Operationele Research 2000De jury heeft de volgende inzendingen ontvangen.
Deelnemer: Lucie Aarts Semidefinite Optimization (SDO) problems form an important class of convex optimization problems. There exist various applications in combinatorial optimization, control and system theory and optimization of engineering structures. For solving SDO problems an important breakthrough was achieved in 1988 by Nesterov and Nemirovsky. They showed that polynomial interior-point methods (IPM's) for solving LO problems were also applicable to SDO problems. Because IPM's seem to provide a powerful approach for solving SDO problems, many results have already been reported in this emerging field. Contrary to the LO case, the search directions in so-called primal-dual methods for SDO are not uniquely defined. This is since the directions that are provided by the original Newton system are not always symmetric. The solution to this problem is to symmetrize the original system. Each strategy to symmetrize the original system results in general in a different Newton system. This report contains the results of an investigation of primal-dual search directions in Semidefinite Optimization. The starting point of my work was the article "On search directions in interior-point methods for semidefinite programming" by M.J. Todd. The goal of my study was to understand the contents of this article and to give more readable proofs of several results about well-definedness, existence and coincidence of search directions. A second goal was to graphically illustrate the search directions. Further I included three other search directions that were proposed after the appearance of Todd's article and I introduced a consistency requirement. A search direction has to fulfil this requirement to be useful. I also looked at the duality theory of Semidefinite Optimization. Deelnemer: Dennis Fok In modern marketing market segmentation is an important topic. It is not possible to consider the market as a homogeneous group of clients anymore. Different clients have different needs. Ideally every client should get a tailor-made treatment. In most cases this is however impossible because of the number and diversity of clients. When focussing on individual clients is not possible, the attention is often directed at different segments in the client population. Therefore much effort is made to segment the market properly. Individual market segments can then be approached differently. A heterogeneous market is divided in (more or less) homogeneous market segments. The research presented in this thesis is an example of an a-priori segmentation approach. In this approach the type and the number of segments is known beforehand. Most of the segmentation bases used by the Rabobank are based on common sense and survey analyses. The risk profile a client has is one of those bases. For example clients investing in high-risk products such as individual stocks form a distinct segment of the total client population. Our research will be mainly based on the relation between the behavior and characteristics of Rabobank clients and their risk profile. Observations on transactions and market values of investments form the behavior. The research is separated in two main parts. In the first part we try to identify factors that explain the risk attitude. Based on the general characteristics of a client we try to predict his/her risk profile. Next to this we identify the most important characteristics that affect the risk profile. It is also interesting to study the behavior of the clients in time. There are of course clients who move between risk categories in time. These particular clients are from a marketing point of view very interesting. Special models similar to Markov models are used to select potential movers. Because most of the clients do not move, some technical problems arise when trying to estimate these models. Only a small percentage of the, very large number of, observed clients move. It is not possible to estimate the model on all observations, since there are simply too many observed clients. We developed a data reduction technique that makes it possible to discard a large portion of the non-movers and use all observed movers in the estimation. Deelnemer: Jan-Willem Goossens Aan de basis van een periodieke dienstregeling, zoals bijvoorbeeld gebruikt door de Nederlandse Spoorwegen, ligt een lijnvoering. Een lijnvoering is een verzamelingen lijnen, waarbij een lijn gezien wordt als een pad in het spoorwegnetwerk. Bij het samenstellen van lijnvoeringen kunnen verschillende doelstellingen gebruikt worden. De klassieke doelstelling in dit verband is het maximaliseren van het aantal reizigers dat rechtstreeks van hun herkomst naar hun bestemming kan reizen. Behalve duidelijke voordelen voor het comfort van de reiziger zal deze doelstelling echter vaak tot een ineffici?te inzet van het schaarse materieel leiden. Dit laatste vormt mede de aanleiding voor het overwegen van een andere, kosten-minimaliserende doelstelling. In deze scriptie komen voor dit probleem ontwikkelde formuleringen en oplossingsmethoden en -technieken aan bod. Afgesloten wordt met een kort overzicht van behaalde resultaten en suggesties voor verder onderzoek. Deelnemer: Ramon Lentink Deelnemer: Vera Raats Een bekend probleem uit de accountantscontrole is het bepalen van een betrouwbaarheidsbovengrens voor de fractie fouten in een populatie posten op basis van een steekproef. Daarbij moet ervan worden uitgegaan dat de inspectie foutloos was. Maar ook accountants (medewerkers) kunnen fouten maken: fouten kunnen over het hoofd worden gezien, aangetroffen 'fouten' kunnen in feite correcte waarden zijn. Om inzicht te krijgen in deze mogelijke beoordelingsfouten wordt uit de steekproef een deelsteekproef opnieuw gecontroleerd - nu door een (foutloos veronderstelde) expert. Op basis van de gecombineerde informatie uit deze twee steekproeven geeft de scriptie schatters voor de fractie fouten in de populatie en voor de kansen op beoordelingsfouten. Verder worden numerieke methoden gegeven voor de berekening van betrouwbaarheidsgrenzen. Ten slotte wordt de Bayesiaanse aanpak besproken, waarbij uitgegaan wordt van bepaalde vooronderstellingen omtrent de foutenkansen. De methoden worden hoofdzakelijk toegelicht aan de hand van een praktijkvoorbeeld, afkomstig uit de sociale verzekeringen. Beoordelingsfouten zullen daar relatief vaak voorkomen, gezien de zeer ingewikkelde aard van de regelgeving op het gebied van de sociale uitkeringen. Merk echter op dat de analyse van toepassing is op algemene kwaliteitscontroleproblemen. Deelnemer: Miranda van Uitert Multi-service networks like ATM and the next generation IP have to deal with a growing number of different services and a large number of users. In the operations of these networks there is clearly a trade off between efficiency and quality of service (QoS). The network provider wants its resources to be used efficiently, whereas the users of the network require a certain QoS level. The admission control algorithm decides whether a new traffic flow can be admitted to the network, taking into account both efficiency and QoS requirements. Current admission control algorithms (i.e., static admission control algorithms based on traffic contract parameters) are very safe in the sense that every user gets the QoS it requests, but are not efficient. A new method to increase the efficiency while maintaining the desired QoS levels, is measurement-based admission control. With measurement-based admission control the admission decision is based on real time measurements of the traffic (or for instance the actual performance) in the network. In this report we investigate several aspects of measurement-based admission control, such as efficiency and safety. First we describe two measurement-based admission control algorithms, an algorithm from the literature and an algorithm we have developed ourselves. Then we use simulation experiments to test the performance of the two algorithms. We use both artificially generated traffic and real time traffic measurements as input for the simulation experiments. Next to the simulation experiments we have developed an analytical model which we will use to gain insight in the sensitivity of measurement-based admission control with respect to traffic characteristics. Finally we present the numerical results we obtained with this model. Based on the results of the simulation experiments we can conclude that measurement-based admission control yields a significantly higher efficiency than static admission control while offering the same QoS. The simulation experiments with real traffic measurements show a utilisation gain of at least 100%. The simulation results and also the analytical model reveal a sensitivity of measurement-based admission control with respect to the traffic characteristics. An important observation is that traffic showing long-range dependent behaviour, in the literature known as `bad traffic', appears to be quite good traffic for measurement-based admission control. We have proved measurement-based admission control to be a very promising method to increase the efficient utilisation of the network resources while offering the appropriate QoS. The study we describe in this report however, is a preliminary investigation of measurement-based admission control, since the current algorithms are not ready for usage in the switches of a network. Hence, it is very important that further research is conducted. Other subjects for further research are the amount of efficiency gain and the robustness of measurement-based admission control.
|
|
23-04-2001 |
