De jury heeft de volgende inzendingen ontvangen.
As with almost any product in any company, cost reduction is a major issue in the roller blinds production process. A large part of the cost of a roller blind is determined by the material used, in particular the tubes and above all the fabrics. Consequently, reduction of the inevitable waste during the production process is favorable. In this thesis we will focus on the one and two-dimensional cutting stock problem within the roller blind production process.
The one and two-dimensional cutting stock problem occur in many industrial processes in many different forms. As a consequence, there has been a lot of attention in literature for both problems throughout the years. However, literature does not offer a solution technique that fits the one-dimensional or two-dimensional cutting stock problem as they occur in the roller blind production process due to their specific requirements.
In this thesis we will propose solution techniques to solve these problems and we will show that using these techniques in practice will result in significant savings in comparison to the current way of working.
This report presents the comparison of two strategies for scheduling patients for the CT in the Erasmus Medical Center Rotterdam. These two strategies are:
In this report, the relationship between economic growth and carbon dioxide emissions is investigated. The hypothesis that this relationship is of the ''Kuznets'' type is explored: do carbon dioxide emissions rise with income until reaching a threshold, after which the relationship becomes negative? It is found that, on a world aggregated level, this hypothesisis indeed supported by our dataset.
The focus is on the validity of parametric (polynomial regression) versus non-parametric (kernel regression) methods of estimation: both are employed and the results are compared. It is concluded that, for this case, the two approaches yield the same result.
Furthermore, the issue of heterogeneity with respect to countries is addressed: can a researcher assume that the shape of the relationship we are studying is the same for different countries? In this report, it is argued that this is not sensible. As an alternative, data aggregation is suggested to investigate the data on a global level.
Nowadays in design optimization, simulation models are often used to determine the characteristics and quality of a product or process. Many of these simulation models unfortunately require much calculation time per simulation run. One way to deal with this problem is to develop a so-called compact model that approximates the simulation model with a function.
Recently, the Centre of Quantitative Methods (CQM) has developed a new method called Optifit which applies simulated annealing to symbolic regression. By not imposing a certain structure on the compact model, it tries to find a compact model that better describes the behavior of the process or product underlying the simulation model.
An important issue in Optifit is controlling the complexity of the compact model. The reasons for this are that complex functions are difficult to understand are more likely to overfit. Overfitted functions are too much conditioned to the training data set and approximate the simulation model poorly for other designs.
The goal of the thesis is twofold. Firstly, to develop a new complexity measure that rewards understandability and penalizes overfitting. Secondly, to find a good method for making the trade-off decision between fit and complexity.
The search for a suitable complexity measure in literature resulted in only one promising measure developed by Vladislavleva. After analyzing this measure and its drawbacks, we introduce our new complexity measure. The new measure is based on the idea, introduced by Vladislavleva, to measure the complexity of a function by the degree of the polynomial necessary to approximate the function with a certain accuracy. The way to determine this approximating polynomial is the main difference with the measure of Vladislavleva. In the new measure, the approximating polynomial is determined with polynomial interpolation through an increasing number of points.
We furthermore discuss Pareto simulated annealing as a technique to make the trade-off decision between fit and complexity.
Finally, we perform experiments on two cases with Pareto simulated annealing combined with the new complexity measure and with the current version of Optifit.
Nonresponse in surveys can be a serious threat to the quality of statistics. Research shows that respondents often differ from non-respondents, so that a bias is created in the results of the survey. In order to correct this, a non-response adjustment has to be made to the outcome of the survey. Giving weights to the respondents is one way of adjusting for the bias. By giving a larger weight to groups underrepresented and a smaller weight to groups overrepresented in the survey response, the bias is corrected for in auxiliary characteristics like age and gender. The construction of these groups, called strata, however, is not straightforward.
The classification tree is a data-mining technique that can be used to construct stratifications in a natural way. A classification algorithm is used to split the sample into strata. In previous research, a new classification algorithm was developed by Statistics Netherlands for constructing such a tree. One disadvantage of this method, however, was the uncertainty about the robustness of the algorithm for different samples.
The main objective of this paper is to find a robust classification method. For this purpose we make use of so-called Random Forests. Random Forests are a group of methods that grow an ensemble of trees by using a random subset of the available information. In our case we base the Random Forest on the proposed classification algorithm.
We investigated two versions of the Random Forest technique. In the first version we randomly select only a part of the available variables at each node of the classification trees. In the second version we randomly select a part of the available sample. We define robustness of a classification method and stability within forests and compare both versions of the random forests to find out if Random Forests are a useful tool for nonresponse adjustment for Statistics Netherlands.
Information and communication technology is becoming part of every day life. Well-known and frequently used applications of ICT are electronic banking and hotel bookings. Much of these kinds of Internet services are even accessible via the current generation mobile phones. For the commercial success of this kind of services, the ability to deliver an acceptable quality of service in terms of the end-to-end response time experienced by the end user is of key importance.
Since delays play a dominant role, we will use the queueing network approach in this thesis.
The mean total sojourn time in a queueing network is a measure for the average end-to-end response time experienced by the client. The variability of the end-to-end response times will be measured by the variance of the total sojourn time in a network. We consider a number of variants of queueing networks. For each network we derive either an exact expression for the mean total sojourn time and an approximation for the variance of the total sojourn time in the network or we develop an explicit, fast-to-evaluate and accurate approximation for the mean total sojourn time in the network.