Online sequential ensembling of predictive fuzzy systems

Abstract Evolving fuzzy systems (EFS) have enjoyed a wide attraction in the community to handle learning from data streams an incremental, single-pass and transparent manner. The main concentration so far lied development of approaches for single EFS models, basically used prediction purposes. Forgetting mechanisms been increase their flexibility, especially purpose adapt quickly changing situations such as drifting distributions. These require forgetting factors steering degree timely out-weighing older learned concepts, whose adequate setting advance or adaptive fashion is not easy fully resolved task. In this paper, we propose new concept streams, which call online sequential ensembling (OS-FS) . It able model recent dependencies on chunk-wise basis: each incoming chunk, trained scratch added ensemble (of before). This induces (i) maximal flexibility terms being apply variable chunk sizes according actual system delay receiving target values (ii) fast reaction possibilities case arising drifts. latter are realized with specific techniques chunks based members over time. We four different variants including various weighting concepts order put higher weights inference certainty during amalgamation predictions final prediction. sense, members, keep mind knowledge about past states, may get dynamically reactivated cyclic drifts, induce dynamic changes process behavior re-occurring time later. Furthermore, integrate properly resolving possible contradictions among similar certainties. onto drifts thus autonomously handled demand fly stage (and adaptation/evolution conventionally done models), yields enormous flexibility. Finally, cope large-scale (theoretically) infinite within reasonable amount time, demonstrate two pruning one atypical high error trends non-diversity members. results showed significantly improved performance compared models better convergence accumulated ahead trends, regular Moreover, more advanced schemes could outperform standard averaging all members’ outputs. contradictory outputs helped improve further. Results wider range application scenarios trend lines well related AI methods OS-ELM MLPs neural networks retrained chunks, slightly worse than on-line bagged (as ensembles), but around 100 times faster processing (achieving low way below requiring milli-seconds samples updates).