Published — International Journal of Computer and Information Technology (ISSN: 2279–0764) A High-Performance Variable Structure Neural Network
Volume 07 — Issue 02, March 2018
Z. Gao, T. Wang, X. Song, C. Wang
Tuzi Network Technology Shanghai
Shanghai, China
E-mail: bottos2050@gmail.com
Abstract — The paper is focused on the analysis and application of a variable structure feed-forward neural network. For issues of segmentation and performance optimization, the coding scheme, operational parameters, initiation of training configuration, generation of sub-population, decomposition of network connection are all factors which should be considered. By showing the main code, the proposed method to achieve a high-performance variable structure neural net is developed. In order to verify its performance, a balancing robot based experiment is conducted.
Introduction — As one of the most successful tools for machine learning and deep learning, neural networks have been applied for various scenarios in fields of image recognition, speech recognition, decision marking, output prediction, process approximation and robot control [1–4]. Although these days the importance of training data, including quality and quantity, has been well realized, the development and performance improvement of intelligent and self-adaptive models are still essential.
The main purpose of this work is to introduce a high-performance variable structure neural network. Segmentation of neural network directly affects its overall efficiency. Following figure shows the horizontal segmentation of a representative feed-forward neural network. It can be found that with this configuration, some connections between nodes are difficult to be divided, for example, the connection between nodes 2 and 3, the connection between nodes 3 and 8, and so on. Since these connections cannot be easily dealt with, the performance evaluations of such kind of segmentation are not achievable.
Therefore, other alternative solutions should be considered. One of the best alternative solutions is to separate the entire network based on dividing the connections layer by layer. In this way, the groups of connections in between layers can be coded with both connectivity and weight. Some methods such as genetic algorithm, swarm algorithm can be considered to deal with the connectivity and weight [5–10].
Read the full publication here:
Join Our Community and Stay Updated!
