The Echo State Network (ESN) is a representative implementation of RC. However, the relatively fixed training scheme and the simple architecture of the standard ESN unavoidably limit improvements of the representation ability and computational ability on many time-series processing tasks such as time-series prediction tasks and time-series prediction classification tasks.
We focus on using Multi-Reservoir Echo State Networks (MRESNs) which is an extended paradigm of the standard ESN. In Ref.[1], we proposed a modular (encoder-decoder) architecture to construct the MRESN based models. Based on this modular architecture, most MRESN-based models can be constructed quickly. We also proposed some extended MRESN-based models with better computational ability on nonlinear time-series prediction tasks than the representative MRESN models [2] [3]. In the future, we will continue on exploring more advanced MRESN-based models for dealing with various time-series processing tasks effectively.
References:
[1] Z. Li and G. Tanaka. "Multi-reservoir echo state networks with sequence resampling for nonlinear time-series prediction." Neurocomputing 467 (2022): 115-129.
[2] Z. Li and G. Tanaka. "Deep Echo State Networks with Multi-Span Features for Nonlinear Time Series Prediction." In 2020 International Joint Conference on Neural Networks (IJCNN), pp. 1-9. IEEE, 2020.
[3] Z. Li and G. Tanaka. "HP-ESN: Echo State Networks Combined with Hodrick-Prescott Filter for Nonlinear Time-Series Prediction." In 2020 International Joint Conference on Neural Networks (IJCNN), pp. 1-9. IEEE, 2020.