Evolutionary multitasking network reconstruction from time series with online parameter estimation

Abstract

Reconstructing the structure of complex networks from time series is beneficial to understanding and controlling the collective dynamics of networked systems. Existing network reconstruction algorithms can only deal with one network reconstruction problem at one time. However, real-world applications typically have multiple network reconstruction tasks and these tasks often relate to each other to certain extent. For the purpose of exploring similar network structure patterns in different tasks, we establish an evolutionary multitasking framework to simultaneously optimize double network reconstruction tasks. In the proposed method, each task is modeled as a single objective problem containing the reconstruction error and the l0-norm of the weight matrix, which takes both the reconstruction error and the structure error into consideration and deals with the NP-hard problem directly. Online parameter learning scheme is employed to learn the parameter automatically controlling the amount of genetic material to exchange, thus avoiding the negative transfer while allowing the useful information pass between tasks. In addition, the least absolute shrinkage and selection operator (LASSO) initialization is employed to further enhance the performance. We apply the evolutionary multitasking framework to reconstruct both synthetic and real networks of evolutionary game, resistor networks, and communication network dynamic models. The experimental results demonstrate that the proposal exhibits competitive performance against state-of-the-arts in terms of all evaluation measures.