TY  - JOUR
A1  - Schneider, Wolfgang
A1  - Scheibler, D.
T1  - Probleme und Möglichkeiten bei der Bewertung von Clusteranalyse-Verfahren: II. Ergebnisse einer Monte-Carlo-Studie
T1  - On the evaluation of clustering algorithms: A Monte Carlo approach
T1  - Probleme et possibilite pour l'evaluation des procedes l'analyse de cluster
N2  - Ziel der vorliegenden Untersuchung war es, Aufschluß über die unterschiedliche Qualität hierarchischer und nicht-hierarchischer (partionierender) Clusteranalyseverfahren zu gewinnen. Die Reproduktionsgüte beider Clusteranalyse-Varianten wurde anhand von 200 Monte-Carlo-Datensätzen (multivariat normalverteilte Mixturen) zu überprüfen versucht, wobei jeweils unterschiedliche Proportionen der Daten-Elemente klassifiZiert werden mußten. Es zeigte sich, daß insgesamt gesehen die hierarchischen Algorithmen nach WARD und LANCE-WILUAMS am besten dazu in der Lage waren, die vorgegebenen Datenstrukturen zu reproduzieren, andererseits aber die herangezogenen partitionierenden KMEANS-Verfahren nicht schlechter abschnitten, wenn die Lösung der WARD-Technik als Start-Partition vorgegeben wurde.
N2  - In this study, a number of hierarchical dustering algorithms and nonhierarchical (i.e. iterative-partitioning) methods were compared with regard to accuracy on the basis of 200 monte carlo data sets. As main results, the two hierarchical procedures by WARD and LANCE-WILUAMS as weil as two nonhierarchicallc-means algorithm using WARDs solution as starting seeds proved tobe most robust. Although some of the remaining algorithms showed acceptabel recovery values when only a certain proportion of the elements had to be classified, it is recommended to choose the few methods mentioned above for particular applications.
N2  - Le but de cette etude est d'obtenir des renseignements sur les differentes qualitees hierarchiques et non-hierarchiques (partionaires) procedes d'analyse de Clusters. Le qualite de reproduction des deux variantes d'analyse de Cluster a ete relevee et controlee a l'aide de 200 groupes de Monte-Cario (multivariation, melange de distribution normale). Pour chacune des proportions differentes, les elements de donnees ont du etre classes. On observe, dans l'ensemble, que l'algorithme hierarchique selon Ward etLance-Williams, est en mesure de reproduire, le mieux, les structures de donnees impliquees. D'autre part, les procedes appliques de KMEANS-partionaires ne se detachent pas pour le moins de ces resultats lorsque Ia solution de Ia technique de Ward a été, au prealablement, donnée comme situation de depart.
KW  - Monte-Carlo-Simulation
KW  - Cluster-Analyse
Y1  - 1983
U6  - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:20-opus-69637
ER  - 
TY  - JOUR
A1  - Chipperfield, Joseph D.
A1  - Dytham, Calvin
A1  - Hovestadt, Thomas
T1  - An Updated Algorithm for the Generation of Neutral Landscapes by Spectral Synthesis
N2  - Background: Patterns that arise from an ecological process can be driven as much from the landscape over which the process is run as it is by some intrinsic properties of the process itself. The disentanglement of these effects is aided if it possible to run models of the process over artificial landscapes with controllable spatial properties. A number of different methods for the generation of so-called ‘neutral landscapes’ have been developed to provide just such a tool. Of these methods, a particular class that simulate fractional Brownian motion have shown particular promise. The existing methods of simulating fractional Brownian motion suffer from a number of problems however: they are often not easily generalisable to an arbitrary number of dimensions and produce outputs that can exhibit some undesirable artefacts. Methodology: We describe here an updated algorithm for the generation of neutral landscapes by fractional Brownian motion that do not display such undesirable properties. Using Monte Carlo simulation we assess the anisotropic properties of landscapes generated using the new algorithm described in this paper and compare it against a popular benchmark algorithm. Conclusion/Significance: The results show that the existing algorithm creates landscapes with values strongly correlated in the diagonal direction and that the new algorithm presented here corrects this artefact. A number of extensions of the algorithm described here are also highlighted: we describe how the algorithm can be employed to generate landscapes that display different properties in different dimensions and how they can be combined with an environmental gradient to produce landscapes that combine environmental variation at the local and macro scales.
KW  - Landschaft
KW  - Monte-Carlo-Simulation
KW  - Brownsche Bewegung
Y1  - 2011
U6  - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:20-opus-68938
ER  -