A Fuzzy Transfer Reservoir Learning Machine Through Domain Enhancement on Multiple Sources

Abstract

While transfer learning through source domain enhancement with the mix-up strategy on multiple sources is applied to reservoir computing (RC) related resource-constrained scenarios, this study aims at addressing two seldom-concerned phenomena: 1) the similarity degrees between the target domain and each of all the source domains may perhaps change after reservoir transformation so as to possibly change their similarity rankings before and after that transformation; 2) the decision boundaries between classes may become more uncertain. In order to achieve this goal, a fuzzy transfer reservoir learning machine (FT-RLM) is proposed based on the well-known leaky integrator echo state network (LI-ESN). In particular, in order to determine which source domains should be enhanced by the mix-up strategy after reservoir transformation, with the theoretical derivation of the mix-up ratios for source domain selection, FT-RLM begins with the use of the mix-up strategy based on the calculated mix-up ratios for source domain enhancement. After that, in order to deal with uncertain decision boundaries between classes, FT-RLM takes the proposed transfer-learning-based fuzzy classifier called parametric-transfer-based Takagi-Sugeno–Kang fuzzy system (TSK-FS) which is trained on both the enhanced source domains and the target domain. Experimental results on real-world datasets validate the effectiveness of the proposed FT-RLM when faced with the above two phenomena in multiple source reservoir transfer learning scenarios under RC-related resource-constrained environments.