Novel Learning Algorithm based on BFE and ABC for Process Neural Network and its Application

In order to improve the generalization capability of process neural network (PNN), a novel learning algorithm is proposed based on basis function expansion (BFE) algorithm and artificial bee colony (ABC) algorithm, named BFE-ABC algorithm. First, the input functions and weight functions are simplified through BFE algorithm. The parameter space is transformed from function space to real number space in this way. Then, the PNN is designed to parametric representation through introducing two Boolean variables and one multidimensional parameter. At last, the multidimensional parameter composed of hidden neurons, expansion items and connection weights is optimized in real number space by ABC algorithm. BFE-ABC algorithm overcomes the premature problem and realizes the global optimization of the structure, connection weights and function expansion form at the same time. It is validated through the prediction experiment of Mackey-Glass chaotic time series. The test results in cylinder head temperature prediction prove the superiority of BFE-ABC algorithm over traditional learning algorithm and the applicability to time-dependent parameter prediction.