Temporally varying classification by a dynamic classifier network is introduced. The dynamic classifier network consists of several independent nonlinear classifiers in parallel. The subclassifiers adapt to the measurements with a variety of adaptation rates. The output of the classifier network can be calculated as a weighted sum of the outputs of each subclassifier. Two methods to optimize the weighting are given. However, even a simple weighting function gives reasonable results. The network might be considered as a temporal associative memory. Because of nonlinearities and the ensuing chaos the behavior of the network can be very complicated. Algorithms to calculate the fractal and correlation dimension are also given. With these dimensions one can estimate how complicated the behavior of a system is and how many parameters are needed to describe its behavior. An extension of geodesic distance transform, called distance transform in curved space, is also presented. This transform can, for example, be used to model dynamic decision manifolds. Some new properties of fractals are also presented. These properties can be utilized efficiently when defining the Lyapunov exponents and the basins of attraction for maps. The methods presented have several application areas.
|Award date||20 Dec 1991|
|Place of Publication||Espoo|
|Publication status||Published - 1991|
|MoE publication type||G5 Doctoral dissertation (article)|
- neural networks
- nonlinear network analysis