A decomposition-based memetic neural architecture search algorithm for univariate time series forecasting

Abstract

Although deep learning has made remarkable progress in time series forecasting, enormous hyperparameters consume a lot of effort to tune. Moreover, to further build the forecasting models with better performance, time series decomposition is usually adopted to mine implicit patterns of the data. Inspired by the time series decomposition, automatically searching for a network architecture after decomposing the time series is proposed. The searching process is non-trivial and has two key challenges: 1) impairment of time series information after decomposing and 2) enlarged search space caused by the huge parameters to be optimized. In this paper, a decomposition-based memetic neural architecture search algorithm is proposed for univariate time series forecasting to address these two challenges. For the first challenge, a general univariate time series forecasting paradigm is designed as the building pipeline of the individual in the proposed algorithm, which considers both the decomposed components and the original series as the compensation information to improve the network representation ability. For the second challenge, with the intrinsic property of representation of individuals in mind, we design a decomposition-based memetic algorithm with a discriminative local search operator to automatically optimize the network configurations. The experimental results on nine benchmarks with four horizons and one application of remaining useful forecasting demonstrate that the discovered architectures by the proposed algorithm achieve competitive performance compared with six methods under aligned settings. Codes and models will be released in https://github.com/EavanLi/dMA-NAS-UTSF.