DERC: Distributed Edge Reservoir Computation for Many Agents

Abstract Reservoir computers have several properties that lend themselves to implementation in resource constrained computational situations. They train quickly, are computationally efficient, and take up very little memory real estate. Also, because reservoir computersimplemented on von Neumann architecture are a simulacrum of aphysical dynamical system, they may be used as an analog to test the ability of new computational substrates before attempting to apply these computational paradigms safety critical situations. We show that the same randomly initialized model may be tuned for and applied to many similar tasks because any training occurs only in the final layer of the architecture. In performing an exhaustive sweep of hyperparameter tunings on our model, we are able to show that the number of parameters which a reservoir is instantiated with is not correlated to its precision in predicting the trajectories of several chaotic ordinary differential equations. We are further able to show that a single reservoir with hyperparameters that have been generally tuned to be in an edge of chaos operating regime are able to perform both similar and disparate computational tasks between many agents. We further show that reservoir computers operating in this regime are able to be applied in real-world computational tasks, including predicting computational load and powergrid utilization with similar accuracy tomore computationally intensive models that would require a much greater computational resource. CCS Concepts • Computer systems organization → Embedded software; Realtime system architecture; • Computing methodologies → Multiagent systems.