Lars Willmes’s research while affiliated with Leiden University and other places

In recent years, the distribution grid planning process has faced the big challenge to integrate renewable energy sources in its planning methodology while preserving a secure and stable provision of electricity. With the currently observable efforts to electrify human mobility all around the world, another new challenge arises for the planning and operation of distribution grids. To address these challenges and to leverage the opportunities that are accompanied by them, new methods for the planning of distribution grids as well as planning decision-supportive approaches and algorithms are needed. The presented approach contributes to the described demands by means of a coupled approach, using both distribution grid time series as well as a genetic algorithm to support decision making in the planning process considering not only new assets for grid reinforcements and extensions but also smart-grid and operational opportunities.

Background: Electronic patient files generate an enormous amount of medical data. These data can be used for research, such as prognostic modelling. Automatisation of statistical prognostication processes allows automatic updating of models when new data is gathered. The increase of power behind an automated prognostic model makes its predictive capability more reliable. Cox Proportional Hazard Regression is most frequently used in prognostication. Automatisation of a Cox model is possible, but we expect the updating-process to be time consuming. A possible solution lies in an alternative modelling technique, called Random Survival Forests. RSF is easily automated and is known to handle the proportionality assumption coherently and automatically. Performance of RSF is not yet tested on a large head-and-neck-oncological-dataset. This study investigates performance of head-and-neck-overall-survival-RSF-models. Performances are compared to a Cox model as golden standard. RSF might be an interesting alternative modelling approach for automatisation when performances are similar. Patients and methods: RSF models were created in R (Cox also in SPSS). Four RSF-splitting-rules were used: log-rank, conservation-of-events, log-rank-score and log-rank-approx. Models were based on historical data of 1371 primary head-and-neck patients, diagnosed between 1981 and 1998. Models contain eight covariates: tumour site, TNM-classification, age, gender, prior malignancies and comorbidity. Model performances were determined by Harrell's concordance error rate, where 33% of the original data served as a validation sample. Results: RSF and Cox models delivered similar error rates. The Cox model performed slightly better (error rate: 0.2826). The log-rank-splitting-approach gave best RSF performance (error rate: 0.2873). According to Cox and RSF, high T-classification, high N-classification and severe comorbidity are very important covariates in the model, while gender, mild comorbidity and a supraglottic larynx tumour are less important. A discrepancy arose regarding importance of M1-classification (see discussion). Conclusion: Both approaches delivered similar error rates. The Cox model gives a clinically understandable output on covariate impact while RSF becomes more of a black box. RSF complements the Cox model by giving more insight and confidence towards relative importance of model covariates. RSF can be recommended as approach-of-choice in automating survival analyses.

Background: Electronic patient files generate an enormous amount of medical data. These data can be used for research, such as prognostic modeling. Automatization of statistical prognostication processes allows automatic updating of models when new data is gathered. The increase of power behind an automated prognostic model makes its predictive capability more reliable. Cox proportional hazard regression is most frequently used in prognostication. Automatization of a Cox model is possible, but we expect the updating process to be time-consuming. A possible solution lies in an alternative modeling technique called random survival forests (RSFs). RSF is easily automated and is known to handle the proportionality assumption coherently and automatically. Performance of RSF has not yet been tested on a large head and neck oncological dataset. This study investigates performance of head and neck overall survival of RSF models. Performances are compared to a Cox model as the "gold standard." RSF might be an interesting alternative modeling approach for automatization when performances are similar. Methods: RSF models were created in R (Cox also in SPSS). Four RSF splitting rules were used: log-rank, conservation of events, log-rank score, and log-rank approximation. Models were based on historical data of 1371 patients with primary head-and-neck cancer, diagnosed between 1981 and 1998. Models contain 8 covariates: tumor site, T classification, N classification, M classification, age, sex, prior malignancies, and comorbidity. Model performances were determined by Harrell's concordance error rate, in which 33% of the original data served as a validation sample. Results: RSF and Cox models delivered similar error rates. The Cox model performed slightly better (error rate, 0.2826). The log-rank splitting approach gave the best RSF performance (error rate, 0.2873). In accord with Cox and RSF models, high T classification, high N classification, and severe comorbidity are very important covariates in the model, whereas sex, mild comorbidity, and a supraglottic larynx tumor are less important. A discrepancy arose regarding the importance of M1 classification (see Discussion). Conclusion: Both approaches delivered similar error rates. The Cox model gives a clinically understandable output on covariate impact, whereas RSF becomes more of a "black box." RSF complements the Cox model by giving more insight and confidence toward relative importance of model covariates. RSF can be recommended as the approach of choice in automating survival analyses.

This study describes first steps taken to bring evolutionary optimization technology from computer simulations to real world experimentation in physics laboratories. The approach taken considers a well understood Laser Pulse Compression problem accessible both to simulation and laboratory experimentation as a test function for variants of Evolution Strategies. The main focus lies on coping with the unavoidable noise present in laboratory experimentation. Results from simulations are compared to previous studies and to laboratory experiments.

Neural networks and Kriging method are compared for constructing fitness approximation models in evolutionary optimization algorithms. The two models are applied in an identical framework to the optimization of a number of well known test functions. In addition, two different ways of training the approximators are evaluated: in one setting the models are built off-line using data from previous optimization runs and in the other setting the models are built online from the data available from the current optimization.

Evolving solutions rather than computing them certainly represents an unconventional programming approach. The general method- ology of evolutionary computation has already been known in computer science since more than 40 years, but their utilization to program other algorithms is a more recent invention. In this paper, we outline the ap- proach by giving an example where evolutionary algorithms serve to program cellular automata by designing rules for their iteration. Three dieren t goals of the cellular automata designed by the evolutionary al- gorithm are outlined, and the evolutionary algorithm indeed discovers rules for the CA which solve these problems ecien tly.

Computer simulations of complex engineering problems have become a standard tool of modern product development and design. The increasing computational power at modest costs leads to a growing interest in directly using computer simulation codes for automatic product optimization. Traditional numerical optimization methods have some drawbacks that make them difficult to use with complex simulation software. Gradient-based methods are always local optimizers, thus requiring additional methods such as random restarts to find global optima. Evolutionary optimization is a way to overcome some of these limitations. This chapter presents a paper that introduces evolution strategies as a robust and fault-tolerant optimization method, which does not rely on gradients, is easily adaptable to massively parallel computing systems and can be used for single and multiple-criteria optimization. It describes a complex and aerodynamical test problem that was solved by an evolution strategy. This paper introduces the basic elements of evolution strategies and addresses their important features such as self adaptation, robustness, and multiprocessor implementations.

Computer simulations of complex engineering problems have become a standard tool of modern product development and design. The increasing computational power at modest costs leads to a growing interest in directly using computer simulation codes for automatic product optimization. Traditional numerical optimization methods have some drawbacks that make them difficult to use with complex simulation software. Gradient-based methods are always local optimizers, thus requiring additional methods such as random restarts to find global optima. Evolutionary optimization is a way to overcome some of these limitations. This chapter presents a paper that introduces evolution strategies as a robust and fault-tolerant optimization method, which does not rely on gradients, is easily adaptable to massively parallel computing systems and can be used for single and multiple-criteria optimization. It describes a complex and aerodynamical test problem that was solved by an evolution strategy. This paper introduces the basic elements of evolution strategies and addresses their important features such as self adaptation, robustness, and multiprocessor implementations.

In this paper we will describe the optimisation of a two-criteria wing-design problem where calculation of the objective function requires the solution of the two-dimensional Navier-Stokes equations. It will be shown that basic concepts of the Strength Pareto Evolutionary Algorithm 2 (SPEA2) and the Non dominated Sorting Genetic Algorithm II (NSGA-II) work well with Evolution Strategies. Results for the wing design problem are presented for the selection operators of SPEA2 and NSGA-II in combination with three different mutation operators. These results are compared with results found by a multi-objective Genetic Algorithm.

A new approach for multi criteria design optimisation is presented in the paper. The problem tackled with this approach is the 2- dimensional design of an aircraft wing. To carry the derandomized step size control also to the multi criteria applications, four different selection schemes are proposed. Furthermore, we present a new method for averaging results of multi objective evolutionary algorithms. This method is then used to compare the results achieved with the proposed algorithms.