PAN RAM Bootstrapping Regressor - A New RAM-Based Architecture for Regression Problems

• Published in: Intelligent Systems Vol. 13654; pp. 574 - 587
• Main Authors: de Souza, Starch Melo, de Lima, Kelly Pereira, da Cunha Carneiro Lins, Anthony José, de Brito, Adriano Fabio Querino, Adeodato, Paulo J. L.
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Bootstrapping; Polynomial approximation; RAM-based neuron; Regression
• ISBN: 3031216881; 9783031216886
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-21689-3_40

RAM-based neural networks have been used a few decades before MultiLayer Perceptrons. Despite their implementability in hardware and speed, they have some drawbacks compared to more recent techniques, particularly related to continuous input variables, which kept them out of mainstream research. About a decade ago, the PAN RAM was an attempt to handle continuous inputs with a polynomial approximator neuron in the n-tuple architecture applied to binary decision problems. Constraints on applications and data preparation still remained. This paper presents an evolution of the PAN RAM neuron that can do regression and a single-layer architecture that is dynamically built based on bootstrapping. The proposed system was benchmarked against the Multilayer Perceptron on three regression problems using the Abalone, White Wine, and California Housing datasets from the UCI repository. In the unicaudal paired t-test carried out on a 10-fold cross-validation comparison measuring the Mean Absolute Error (MAE), the proposed system performed better than the Multilayer Perceptron (MLP) on the abalone and white wine datasets. In contrast, for the California dataset there was no significant improvement, all at a 0.05 significance level.