Training set generation using fuzzy logic and dynamic chromosome based Genetic Algorithms for plant identifiers

Training set is one of the main critical sections in Neural Network, generating of it with prior knowledge can be extremely efficient. In this paper we have tried to explore the potential of using previously generated training set (not randomly) for the training of Dynamic Neural Network. The neural network was used as the core of identifier which tries to identify the internal behavior of structure-unknown non-linear time variant dynamic system. In this work, we use genetic algorithm (GA) with dynamic length of chromosomes for generating different training sets based on fuzzy logic ranking system which used as the fitness function of GA. In this regard we extract some features from each training set, in frequency and time (stochastic) domain and consequently set a rank for each. We use industrial robot manipulator for the case study, because of its fully dynamical behavior. The manipulator is simulated with professional simulation software (consist of Solidwork, Visual Nastran 4D and Matlab/Simulink). It is shown that: by using this approach, the error rate of modeling has been decreased and therefore the identifier performance and resolution increase to the levels which gained by using fully random generated signals as training set.