Estimating and controlling dynamical systems through coordinated spikes Spike Coding Networks for Kalman filtering and LQR control

Slijkhuis, Filip

Estimating and controlling dynamical systems through coordinated spikes Spike Coding Networks for Kalman filtering and LQR control

Files

Slijkhuis, F. s-1032106-MSc-MKI92-Thesis-2022.pdf (1.3 MB)

Authors

Slijkhuis, Filip

Issue Date

2022-05-03

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16355

Abstract

Biological neural networks are able to control animal bodies robustly and efficiently, through highly sparse and irregular spiking patterns. In contrast, current implementations of control via artificial spiking neural networks often do not adhere to the spiking patterns found in the brain. In the neuroscience theory of Spike Coding Networks (SCNs), irregularity, sparsity, and robustness naturally arise as a consequence of coordinated spiking across a neural network, but it is not clear how to use SCNs for control. Here, we build upon the SCN framework and describe an analytical solution for estimation and control of dynamical systems with spiking networks, without the need for neural learning and optimization. The resulting networks work as spiking equivalents to a state estimator (Kalman filter) and an optimal controller (Linear–quadratic regulator). Results on numerical experiments, using the spring-mass-damper and cartpole systems, show that the networks retain their irregular and sparse spiking patterns, whilst being robust against neural death. The framework offers a powerful perspective on how precise control can be achieved through biological and artificial neural networks, and opens the way for deploying fast and efficient neuromorphic controllers. Keywords: Spiking Neural Networks · State Estimation · Optimal Control · Dynamical Systems.