Novel direct and self-regulating approaches to determine optimum growing multi-experts network structure

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Growing Multi-Experts Networks (GMN) offer a powerful framework for complex modeling.
• Existing GMN approaches can be complex due to numerous control parameters.
• Optimizing GMN structure is crucial for efficient and effective performance.

Purpose of the Study:

• To present two novel approaches for determining the optimal structure of Growing Multi-Experts Networks (GMN).
• To introduce a Self-Regulating GMN (SGMN) algorithm to address the ergonomic limitations of traditional GMN.
• To evaluate the performance of GMN and SGMN against other neural networks and statistical models.

Main Methods:

• The direct method utilizes expertise domain, levels, and local expert clustering via Growing Neural Gas (GNG).
• Error distribution is employed to apportion errors among local experts, followed by a redundant expert removal algorithm.
• The proposed SGMN incorporates self-adaptive learning rates and a modified fully self-organized simplified adaptive resonance theory for clustering.

Main Results:

• SGMN demonstrates comparative or superior performance to GMN across four benchmark examples.
• SGMN exhibits reduced sensitivity to learning parameter settings compared to GMN.
• Both GMN and SGMN outperform existing neural networks and statistical models in benchmark tests.

Conclusions:

• SGMN provides a more ergonomic and robust alternative to traditional GMN.
• The proposed methods, particularly SGMN, show significant potential for building novel nonlinear models from local linear models.
• SGMN's efficacy is validated through industrial applications and a control problem, demonstrating consistent results.