Estimating local approximation accuracy with the network of hybrid neuron-like units
Сергей Витальевич Попов, Кристина Александровна Шкуро · 2013
Accuracy is one of the most important properties of the solution to any practical problem. Neuro-fuzzy networks usually generate point estimates of the process under consideration, and the accuracy is estimated on average for the whole dataset. This is the easiest way of accuracy estimation and it is justified for most cases, however it is not enough in some situations, where approximation accuracy may be clearly non-uniform across the dataset. In this paper, a network of hybrid neuron-like units is considered, which is expanded to deliver local accuracy estimates. The architecture is constrained by a priori information about the properties of the input signals and the system being modeled and is subsequently optimized on a synaptic level by an evolutionary algorithm. Introduction of a priori information into evolutionary process enables a gray-box approach to systems modeling. Local accuracy estimation provides vital information for subsequent decision making and increases method’s value for the users. The proposed approach is quite general and can be applied to many popular neural networks, e.g. MLP, FIR networks or any other neural and neuro-fuzzy networks (including emerging ones) that are special cases of the network of hybrid neuron-like units.