Comparison of Node Insertion Algorithms for DelaunayNetworks

This paper compares three diierent node insertion strategies which can be used to incrementally construct Delaunay triangulation models. Delaunay networks may be used to eeciently store low dimensional nonlinear models, and can therefore be used in a wide range of real-time applications. However, there are no direct node selection methods, and it can be shown that the network's gener-alisation abilities are strongly aaected by the triangular partitioning of the input space. The three iterative, constructive node insertion algorithms (maximum error, local weighted error and one-step-ahead optimal search) are compared using two data sets, and conclusions are drawn about the quality of the extracted triangulation and the algorithms' computational costs.