Today the centers are obtained by K-means clustering. This leaves no basis functions on the edges of the training data domain, which either lead to large radius (if automatically tuned) or slightly larger errors on the boundary. Consider alternative ways to obtain a good spread of centers on the domain, that also includes some points on the boundary of the domain.
Today the centers are obtained by K-means clustering. This leaves no basis functions on the edges of the training data domain, which either lead to large radius (if automatically tuned) or slightly larger errors on the boundary. Consider alternative ways to obtain a good spread of centers on the domain, that also includes some points on the boundary of the domain.