Adaptive evaluation of complex time series using nonconventional neural units

This paper introduces new adaptive methodology for monitoring of variability (level of chaos) of complex time series by utilization of cognitive capabilities of nonconventional neural architectures. Real-time sample-by-sample evaluation of complex system behavior is based on monitoring of adaptable parameters of neural architectures during their adaptation. The level and changes of complexity of system behavior are adaptively monitored in real time and can be stored for further evaluation. The proposed technique performs sensitivity-scalable sample-by-sample monitoring and variability change detection of system behavior that is achieved as the system output behavior is transformed to approximated system parameter space by the adaptation of a special forced higher-order nonlinear neural unit. The nonconventional neural unit is implemented as an adaptive forced nonlinear dynamic oscillators, i.e., with adaptable forcing periodic inputs. Adding forcing adaptable inputs increases the approximating capability of neural architecture; the forcing adaptable neural inputs are initially configured upon analysis of frequency spectra of the evaluated time series. It is demonstrated that monitoring of system parameters during the adaptation of forced dynamic neural architecture can reveal important attributes of complex system behavior in real time, and it is capable of sensitive both instantaneous and long-term monitoring of changes of chaotic system behavior. In principle, the proposed methodology is universal and is not limited to evaluation of only time series and not only by nonconventional neural units. Simulation results on deterministic, however, highly chaotic data are shown to explain the new methodology and to demonstrate its capability to reflect the level of chaos in a signal and to detect small changes of chaos in a signal.