Use of modular architectures for time series prediction

摘要

Recently, a lot of papers have been published in the field of time series prediction using connectionist models. Nevertheless we think that one of the major problem with is rarely treated in the literature is related to the choice of input parameters (embedding dimension and delay). In this paper, we propose two modular approaches to this problem and apply them to a sunspot-related time series. Experimental results are then compared to a single multi-layer perceptron in order to estimate performances of these models.