A mixture-of-experts framework for adaptive Kalman filtering

Area of Science:

• Engineering
• Computer Science

Background:

• Kalman filtering is a widely used technique for state estimation.
• Traditional Kalman filters assume known system parameters, limiting their performance in uncertain environments.
• Adaptive filtering techniques are needed to handle time-varying or unknown system dynamics.

Purpose of the Study:

• To propose a novel modular and flexible adaptive Kalman filtering approach.
• To enhance estimation accuracy, responsiveness, and computational efficiency compared to existing methods.
• To introduce on-line adaptation mechanisms for filter parameters.

Main Methods:

• A mixture-of-experts framework regulated by a gating network.
• Each expert is a Kalman filter with different parameter realizations.
• Gating network adapts expert weights based on performance.
• Periodic enhancement using recursive quadratic programming or genetic algorithms for parameter adaptation.

Main Results:

• The proposed filter bank demonstrates superior estimation accuracy.
• It exhibits a quicker response to changing environments.
• It offers improved numerical stability and computational efficiency over classical methods.
• Real-time implementation is feasible with the proposed parameter adaptation schemes.

Conclusions:

• The proposed adaptive Kalman filtering approach offers significant advantages.
• It provides a robust and efficient solution for state estimation in uncertain and dynamic systems.
• The modular design and adaptive mechanisms enhance its applicability across various domains.