Adaptive Evolution and Deployment of Neuromorphic Agents

Neuromorphic computing systems are attractive for real-time control at the edge because of their low power operation, real-time processing capabilities and their potential for rapid online learning. In this work, we describe NODES, a framework for performing real-time evolution of spiking neural networks for neuromorphic systems at the edge. This framework enables the generation and deployment of networks specializing in solving the desired application as it exists under the current real-world conditions by reverse-engineering those conditions and replicating them in simulation for training. We apply this approach to a pole-balancing application as well as a real-time combustion engine control application and examine the exhibited behaviors and what they indicate about the capabilities and limitations of this framework.