Efficient Brain-Inspired Control Strategies: Cloud-Based Online Supervised Learning for Spiking Neural Networks in Dynamical Systems

This study addresses the challenges of implementing control policies in dynamical systems using brain-inspired computing tools, focusing on efficiency, reliability, and robustness. Spiking Neural Networks (SNNs), inspired by biological neural circuits, are known for their efficiency and scalability. However, the lack of efficient local learning rules hinders their application in online learning for dynamical systems control. The study proposes a cloud-based supervised online training strategy for recurrent SNNs in the control of dynamical systems. SNN communicates with a cloud-based model-based control system, updating its weights using a biologically plausible learning strategy until a desired threshold is reached. Notably, SNN learns the control signal independently of underlying dynamics and implemented controllers, making it versatile for arbitrary plants and control methods. Simulation results demonstrate acceptable performance in terms of network convergence and control accuracy. The proposed strategy, with only five neurons, achieves control tasks with constrained weight updates, using approximately 5% of possible spikes compared to traditional neural networks, without a significant impact on control accuracy.