Using Taguchi's method of experimental design to control errors in layered perceptrons

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Artificial neural networks (ANNs) are widely used for function approximation.
• A major challenge is minimizing errors when ANNs are deployed in real-world scenarios.
• Errors can stem from noisy or unrepresentative training data, inappropriate network flexibility, and noisy input data during operation.

Purpose of the Study:

• To experimentally rank the significance of different error sources in ANNs.
• To develop design principles for reducing error magnitude and variance in deployed ANNs.
• To improve the reliability and accuracy of ANNs in practical applications.

Main Methods:

• Conducted experiments to evaluate various sources of error in artificial neural networks.
• Analyzed the impact of training data quality (noise, sampling rates) on network performance.
• Assessed the influence of network architecture flexibility and input data noise on error levels.

Main Results:

• Quantified the relative contribution of training data representation, network flexibility, and input noise to overall network error.
• Identified specific design parameters that significantly impact error magnitude and variance.
• Established a hierarchy of error sources based on experimental findings.

Conclusions:

• Understanding the primary drivers of error is crucial for effective ANN design.
• Implementing specific design principles can substantially mitigate errors in deployed ANNs.
• This research offers practical guidance for building more robust and accurate function approximation networks.