Ferdaus, M. M., Al-Mahasneh, A. J., Anavatti, S. G., Senthilnath, J. (2024). A compact meta-learned neuro-fuzzy technique for noise-robust nonlinear control. Applied Soft Computing, 166, 112149. https://doi.org/10.1016/j.asoc.2024.112149
Abstract:
Neuro-fuzzy systems show promise for adaptive control but can become complex due to the need to learn many parameters. This paper presents a resilient nonlinear controller that combines a simplified neuro-fuzzy system (Simp NFS) and simplified neural network (Simp NN) with only two meta-learnable parameters. This architecture enables fast and stable adaptation in uncertain nonlinear discrete-time systems. Simp NFS utilizes interpretable hyperplane-based rules without antecedent parameters, simplifying the learning process to consequent weights. Simp NN reduces complexity by replacing hidden-output weights with their mean. The hybrid auto-adaptive controller (HAC) combines the advantages of Simp NFS and Simp NN, significantly reducing the number of adaptive parameters compared to standard neuro-fuzzy methods for real-time control with limited resources. Simp NFS provides structural adaptivity to handle uncertainties,
while Simp NN ensures reliable disturbance attenuation. The stability of HAC is proven using Lyapunov analysis. Extensive testing on challenging single-input single-output (SISO) and multi-input multi-output systems (MIMO) demonstrates that HAC improves performance by up to 82.55% compared to existing techniques. Key innovations include an ultra-compact metalearned architecture, transparent online evolution of hyperplane clusters, and enhanced modeling capability for nonlinear uncertain systems. This interpretable neuro-fuzzy approach could enhance autonomy and safety by maintaining model transparency.
License type:
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Funding Info:
There was no specific funding for the research done