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

Abstract

This work proposes a novel hybrid Adaptive Particle Swarm Optimization-Back-propagation algorithm for training feed-forward neural networks to identify nonlinear dynamical systems. The approach begins by using Particle Swarm Optimization to optimize the network weights, followed by back propagation to fine-tune the optimized weights, thereby improving the overall solution quality. To prevent early convergence, Particle Swarm Optimization parameters such as inertia weight and other hyperparameters are dynamically adjusted based on a performance index, which is calculated as the difference between the fitness value of the global best solution across consecutive iterations. Convergence analysis using Lyapunov stability theory is also conducted to ensure the proposed algorithm converges to a stable solution. The proposed hybrid approach is evaluated on three benchmark nonlinear problems to validate its effectiveness. Experimental results demonstrate that the hybrid algorithm outperforms traditional Particle Swarm Optimization and back-propagation algorithms in terms of convergence, accuracy, and robustness.