Training recurrent neurocontrollers for robustness with derivative-free Kalman filter

Area of Science:

• Control Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Training neurocontrollers for robust performance on discrete-time physical system models is crucial.
• Recurrent neural networks (RNNs) are employed as neurocontrollers, requiring effective training methods for systems with parameter and signal uncertainties.

Purpose of the Study:

• To develop a robust training methodology for recurrent neurocontrollers applied to discrete-time physical systems with known parameter distributions.
• To extend model reference control capabilities for enhanced neurocontroller performance and applicability.

Main Methods:

• Neurocontrollers implemented as RNNs are trained by evaluating output sensitivities to weight perturbations.
• A derivative-free Kalman filter algorithm, extended by Feldkamp et al., is utilized for neural network training.
• Training minimizes a quadratic cost function averaged over diverse system models, employing a model reference control approach.

Main Results:

• The training process yields a robust recurrent neurocontroller with fixed weights, ready for deployment.
• The derivative-free Kalman filter training method combines second-order effectiveness with applicability to both differentiable and non-differentiable systems.
• The model reference control approach significantly extends previous capabilities, demonstrated through two illustrative examples.

Conclusions:

• The proposed training methodology effectively produces robust recurrent neurocontrollers for uncertain discrete-time physical systems.
• The integration of a derivative-free Kalman filter algorithm offers a versatile and powerful tool for neural network training in control applications.
• This work advances the field of neurocontroller design by providing a robust and broadly applicable training framework.