Model Inference of a Dynamic System by Fuzzy Learning of Geometric Structures

Abstract

One of difficult tasks on dynamic systems is the exploration of connection models of variables from time series data. Reasonable time regions for constructing the models are crucial to avoid improper models or the loss of important information. We propose fuzzy learning of geometric structures to find reasonable time regions and proper models to reveal varying laws of system. By comparing values of fuzzy merging function for shorter time regions and fuzzy unmerging function for larger varying actions, reasonable model regions are inferred. Experimental results (for both simulated and real data) show that the proposed method is very effective in finding connection models adaptive to the evolution of a dynamic system, and it detected large varying actions in the regions below preset minimal region length, whereas the non-fuzzy learning method failed.