Abstract: A novel principium and topology of modeling and optimizing is introduced, which employs like-neural networks based on system architecture. Known from traditional artificial neural networks modeling, existing knowledge about the investigated object is taken into account fully in the novel method, hence physical meanings can be found from the weights of the network. In the training phase, optimized architecture parameters of non-linear system can be obtained by adjusting the weights of networks. The feasibility and validity of the method is illustrated by modeling and optimization of direct-acting relief valve. A new method is provided for non-linear systems modeling and optimizing.

Key words: Like-neural networks, Modeling, Optimization