A Target Value Control While Training the Perceptrons in Changing Environments

Abstract

To ensure fast adaptation and security of social and computerized systems to changing environments, targets of perceptron based classifiers ought to vary during training process. To determine optimal differences between target values (stimulation, arousal) we suggest using genetically evolving multi-agent systems aimed to extract necessary information from sequences of the changes. A specially designed additional feedback chain allows updating the target values faster.