A feature-recombinant asynchronous deep reservoir computing for modeling time series data

Modeling time series data is an important issue in many areas. Reservoir computing (RC) is a promising tool to build time series model due to its dynamic characteristic and simple training way. Asynchronous deep reservoir computing (ADRC) is an improvement version of the traditional RC. It solves time-dependent tasks more efficiently than traditional RC because of its rich dynamics and flexible short-term memory (STM). Nevertheless, it has been an open issue to design RC's or ADRC's reservoir topology owing to large amounts of random factors. To promote the solution of this problem, the paper proposes a feature-recombinant ADRC (FR-ADRC) for modeling time series data. In the FR-ADRC scheme, the first sub-reservoir is designed as the feature-adaptive layer, and a trainable matrix C is introduced into this layer. By learning C, the singular value (SV) distribution of the first layer could be adjusted. Further, the principal components of the reservoir topology can be extracted by the principal component analysis (PCA). Then a new temporary reservoir is constructed by recombining these extracted components. The subsequent information processing is carried out based on the recombinant reservoir, which can be adaptive to the input signals. The validity of the FR-ADRC is tested by modeling some numerical and real-life time series data. Experimental results show that the proposed approach is better than the traditional ADRC in modeling precision, generalization ability and stability.