Shkuro K. Training methods for neuro-fuzzy networks with specialized architectures

An architecture synthesis method for the network based on hybrid neuron-like units is proposed, including the architecture encoding method and an evolutionary architecture optimization method. A special template is initialized on the basis of a priori information about the properties of the input signals and the underlying system. This template then limits the structural parameters during the evolutionary process. Additional free parameters provide a tradeoff between local and global priorities of the evolutionary search. A novel network architecture based on hybrid neuron-like units for local accuracy estimation is proposed that contains a separate accuracy estimation block. The modified psi-transform method to tune networks based on hybrid neuron-like units is proposed featuring the inequalities restrictions to limit the search space and an improved gravitational search method to refine the global extremum coordinates. A synthesis method for networks based on hybrid neuron-like units is further developed featuring a two-stage structural and parametric optimization of the network. The effectiveness of the proposed methods is proved experimentally on test and real-world data.