A Novel Estimation of Tree Length Using Neural Network Approaches in Phylogenetic Analysis

Abstract

     The article studies the integration of artificial neural networks (ANNs) in the phylogenetic analysis of scriptinformatics, focusing on the historical evolution of Arabic, Aramaic, and Middle Iranian scripts. By treating scripts as taxa, pattern systems are exploited to delineate evolutionary trajectories, utilizing an optimised feature selection method and neural network architectures fitnet and feedforwardnet. The presented approach leverages publicly accessible genetic sequence datasets and applies comprehensive preprocessing, including a novel feature extraction process, normalisation, and cross-validation, to predict phylogenetic tree lengths. The results indicate a superior performance of the feedforward neural network, particularly with an architecture of 16 nodes in the first layer and 6 in the second. In machine learning hyperparameters such as learning rate and network size control the training process and affect model performance. Reduced computational time is significantly achieved through meticulous optimisation of hyperparameters without exhaustive phylogenetic analysis. The cross-validation process underlines the model’s robust predictive capacity, paving the way for advanced computational tools in scriptinformatics and evolutionary studies.