In this paper an extension of the established training algorithm for nonlinear system identification called
is presented . It is a heuristic tree-construction method that trains a local linear neuro-fuzzy network. Due to its very simple partitioning strategy,
is a fast and robust modeling approach, but has a limited flexibility. Therefore a new merging approach for regression tasks is presented, that can rearrange the local model structure in the input space, without harming the global model complexity.