Reservoir consisting of diverse dynamical behaviors and its application in time series classification

Time series classification (TSC) has been tackled through a wide range of algorithms. Seminal reservoir computing (s-RC) is composed of a recurrent neural network with random parameters serves as a dynamical memory and is a well-known end-to-end neural networks (NNs) that have been applied to TSC problems. Although the s-RC architecture is suited for dynamic (temporal) data processing, nevertheless choosing the proper design of reservoir plays a crucial role in the efficiency of reservoir computing (RC) in comparison to state-of-the-art fully trainable NNs. In contrast with a large body of researches that has been focused on the aspect of sparsity in the design of RC, in this article, the role of a variety of dynamical behaviors (e.g. stability, periodicity, high-periodicity, and chaos) in RC design is empirically investigated in terms of the richness of the developed dynamical (temporal) representations. Finally, it is shown that RCs with rich dynamical behaviors outperform RCs with a limited spectrum of dynamic behavior in TSC tasks. To evaluate the TSC adaptability of the newly proposed RC framework and state-of-the-art NN-based methods, different experiments on 15 multivariate time series datasets (UCR and UEA datasets) were performed. Our findings divulge that the proposed framework outperforms other RC methods in learning capacity and accuracy and attains classification accuracy comparable with the best fully trainable deep neural networks.