Strategies for improving neuro-fuzzy identification of nonlinear dynamical systems

The performance of neuro-fuzzy schemes strictly depends on the informative potential about the system dynamics captured by the fuzzy rule base used to build the functional relationship between the input and output system variables, as well as on the proper initialization of the estimation procedure and on the adopted optimization algorithm. This paper concentrates on aspects connected with both the construction of a meaningful fuzzy rule base and the adaptation of the learning rate in the back-propagation algorithm with the goal to build an efficient and robust simulator of the dynamics of complex nonlinear systems. The key idea of our approach consists in the integration of qualitative modeling methods with fuzzy systems. The fuzzy model is derived from rules which express the transition from one state to the next one. Such rules are automatically built by encoding the qualitative state descriptions of the system dynamical behaviors inferred by the simulation of the qualitative model. The resulting hybrid method allows us to initialize properly both the neuro-fuzzy identifier and its parameters. The learning process is further improved by choosing the learning rate within convergence bounds