Feedback-based optimization of feed-forward neural network for the modeling of complex nonlinear dynamical systems...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Nonlinear dynamical systems identification is crucial in various scientific fields.
• Training feed-forward neural networks often faces challenges like slow convergence and local optima.
• Existing optimization algorithms may not sufficiently address the complexities of nonlinear system identification.

Purpose of the Study:

• To propose a novel hybrid Adaptive Particle Swarm Optimization-Back-propagation (APSO-BP) algorithm.
• To enhance the training efficiency and accuracy of feed-forward neural networks for nonlinear dynamical systems identification.
• To ensure algorithm stability and robustness through dynamic parameter adjustment and convergence analysis.

Main Methods:

• A hybrid approach combining Particle Swarm Optimization (PSO) for initial weight optimization and Back-propagation (BP) for fine-tuning.
• Dynamic adjustment of PSO parameters (e.g., inertia weight) based on a performance index to prevent premature convergence.
• Convergence analysis using Lyapunov stability theory to guarantee a stable solution.

Main Results:

• The proposed APSO-BP algorithm demonstrated superior performance over traditional PSO and BP methods.
• Experimental validation on three benchmark nonlinear problems confirmed the algorithm's effectiveness.
• The hybrid approach achieved better convergence speed, higher accuracy, and improved robustness.

Conclusions:

• The novel hybrid APSO-BP algorithm offers a significant advancement in training neural networks for nonlinear system identification.
• Dynamic parameter adaptation is effective in overcoming premature convergence issues in PSO.
• The algorithm provides a stable, accurate, and robust solution for complex nonlinear system modeling.