Structure evolution for time-delay neural networks

• Published in: 9th International Conference on Artificial Neural Networks: ICANN '99 pp. 667 - 672
• Main Author: Sick, B
• Format: Conference Proceeding Journal Article
• Language: English
• Published: London IEE 1999
• Subjects: Neural nets; Optimisation techniques; multilayer perceptrons; rank-based selection; delays; time-delay neural networks; evolutionary algorithms; graphical coding; genetic algorithms; hidden neurons; feedforward neural nets
• ISBN: 0852967217; 9780852967218
• ISSN: 0537-9989
• DOI: 10.1049/cp:19991187

The paper presents a structure finding algorithm for time-delay neural networks based on the working principle of evolutionary algorithms. Multilayer perceptrons, which are a subclass of time-delay neural networks, can also be constructed. The algorithm selects appropriate input features for the neural networks from a set of possible inputs, finds optimal values for the number of layers and hidden neurons, constructs connections between neurons, and determines the ideal values of time-delays. The approach uses a new, graphical coding scheme, a rank-based selection mechanism, and seventeen reproduction operators for mutation and crossover. The advantages of this approach are shown by means of an application example (tool wear estimation in turning).