The analysis of the addition of stochasticity to a neural tree classifier

Abstract

This paper describes various mechanisms for adding stochasticity to a dynamic hierarchical neural clusterer. Such a network grows a tree-structured neural classifier dynamically in response to the unlabelled data with which it is presented. Experiments are undertaken to evaluate the effects of this addition of stochasticity. These tests were carried out using two sets of internal parameters, that define the characteristics of the neural clusterer. A Genetic Algorithm using appropriate cluster criterion measures in its fitness function was used to search the parameter space for these instantiations. It was found that the addition of non-determinism produced more reliable clustering performances especially on unseen real world data. Finally, deliberately changing the tree shape by varying key parameters was investigated, illustrated and systematically analysed.

Optimising a stochastic dynamic neural tree

“…The addition of randomness to a hierarchical neural tree clusterer have some benefit in helping the model avoid local minima in its implicit cost function. This approach is well known in the field of optimisation [9].…”

“…A more recent DNTN is the Stochastic Competitive Evolutionary Neural Tree (SCENT), [8,9]. This model is comparatively robust, with respect to its parameter settings when compared with other DNTNs [8].…”