Can adaptive diffusion networks do better with less data?
Daniel Gilio Tiglea. Can adaptive diffusion networks do better with less data? Doctoral Thesis, Electrical Engineering, University of São Paulo, 2024
Abstract
Adaptive diffusion networks consist of a collection of agents that can measure and process locally streaming data, and that can cooperate with one another to improve the overall performance. Since their inception, these networks have consolidated themselves as interesting tools for distributed estimation and learning, and have spun several types of solutions for these problems. To reduce the amount of data measured, processed, and transmitted over these networks, several techniques have been proposed in the literature, which usually affect the performance of the original solutions, but are necessary in order to extend the network lifetime. In this work, in addition to an extensive literature review, we present sampling techniques that eliminate the need to measure and process the data at every node and every time instant. By controlling the sampling of the nodes based on their estimation error, the proposed techniques are able to maintain the convergence rate of the original solutions, while achieving a lower computational cost and better performance in the steady state. This comes at the expense of only a slight increase in the computational cost during the transient phase in comparison with that of the original solutions. Moreover, with slight modifications, the techniques presented can also be used to restrict the number of transmissions between the nodes in the network. Lastly, we conduct a theoretical analysis in order to understand the performance of the proposed solutions, which agrees with the simulation results.