Neural network learning of optimal Kalman prediction and control.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Control Theory

Background:

• Optimal estimation and control methods, like Kalman filtering, exist for linear systems.
• Neural network (NN) algorithms for prediction and control are abundant, but none learn Kalman prediction or control.
• The special case of stationary control is the only exception for NN-based Kalman control.

Purpose of the Study:

• To demonstrate a recurrent neural network (NN) capable of learning optimal Kalman prediction and control (KPC).
• To show that system identification can also be learned by this NN.
• To explore the architectural similarities between the developed NN and the mammalian cerebral cortex.

Main Methods:

• A recurrent neural network (NN) composed of linear-response nodes was designed.
• The NN was trained using only a stream of noisy measurement data.
• The NN architecture was analyzed for constraints imposed by Kalman prediction and control (KPC) requirements.

Main Results:

• The developed NN successfully learned optimal Kalman prediction and control (KPC) from noisy data.
• System identification was also achieved using the same NN.
• The NN architecture exhibited significant resemblances to the local-circuit architecture of the mammalian cerebral cortex.

Conclusions:

• The study presents the first NN algorithm capable of learning Kalman prediction and control (KPC).
• The architectural parallels suggest that cortical circuits might perform prediction, estimation, and control functions.
• Further research is needed to fully understand the biological implications of these findings.