Resumen |
This article presents a Dendrite Morphological Neuron model learned by Linkage Trees (LT-DMN). It is presented as an alternative to modern DMN model training approaches based on k-means clustering that must tune the number of dendrites per class by defining a k-value. Also, the k-means based methods have a problem of non-reproducibility and, for each potential solution, they may present the risk of falling into local minima. The LT-DMN algorithm selects the centroids from a deterministic hierarchical clustering, which builds a linkage tree for each class of patterns. In addition, the simulated annealing algorithm is used to automatically fit a suitable cut-off point in the structure of each tree that minimizes the classification error and the number of dendrites. The proposed method is evaluated on nine synthetic data sets and 17 real-world problems. The results reveal that the proposed method is competitive or even better than seven DMN models from the literature. Furthermore, LT-DMN achieves low architectural complexity by using few dendrites. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG. |