AQPSO-based Self-organization Learning Of RBF Neural Network

As a feedforward neural network,RBF?Radial Basis Function?neural network has been widely used in the fields of model prediction,intelligent control and pattern recognition.However,the topology of the conventional RBF neural network can not be adjusted in real time according to the specific problems,which greatly restricts the application of RBF neural network.Therefore,aiming at the structure design and parameter optimization of RBF neural network,taking the prediction of traffic flow as the application background and combining the global search ability of Quantum Particle Swarm Optimization?QPSO?algorithm,a self-organizing learning method of RBF neural network based on Adaptive QPSO?AQPSO?is proposed.Firstly,the AQPSO algorithm is designed to overcome the disadvantage that the original QPSO algorithm is easy to fall into the local optimal.Secondly,the proposed AQPSO algorithm is applied to the parameter learning.Moreover,the rules of RBF neural network structure adjustment are designed to realize self-organizing learning of RBF network structure and parameters.Finally,the effectiveness of the proposed method is proved by the experimental simulation of nonlinear system identification and short-term traffic flow prediction.The specific research contents are as follows:1.An improved quantum particle swarm optimization algorithm,AQPSO algorithm,is proposed for the QPSO algorithm to be easily trapped in the local optimal solution.Firstly,the weight coefficient?_i,j?t?of particles is introduced to design the best location m_??t?of the average value of weights and replace the m?t?of the original algorithm to evaluateL_i,j?t?.The adaptive shrinkage expansion coefficient?_i,j?t?is designed to improve the global search ability of particles.Then,by establishing the AQPSO Markov model,the convergence of the AQPSO algorithm is analyzed.Finally,the validity of AQPSO is proved by four benchmark functions.2.According to the parameter learning problem of RBF neural network,the proposed AQPSO algorithm is applied to the parameter learning.First of all,the parameters?center,expansion constant,the output weights?of RBF neural network to form a multi-dimensional vector,as AQPSO algorithm of particle optimization,in order to obtain parameters within the scope of global optimal solution space.Then,the RBF neural network optimized by AQPSO algorithm is applied to the prediction of Mackey-Glass chaotic time series.The results show that compared with the traditional PSO algorithm and the QPSO algorithm,it is better to use the proposed algorithm to learn the parameters of RBF neural network.3.The dynamic optimization of RBF neural network structure is an effective method to ensure that the RBF neural network is always working in the appropriate structure state.In order to obtain an effective method for dynamic adjustment of it,the RBF structure is adjusted according to the change of the task.Based on the in-depth analysis of the existing research results,a self-organizing learning method of RBF neural network based on AQPSO is proposed,which realizes the simultaneous optimization design of RBF network structure and parameters.Moreover,the convergence of AQPSO-SORBF neural network is analyzed in depth.Finally,through the simulation research of nonlinear system identification and short-term traffic flow prediction,it is proved that the proposed method can not only achieve better learning performance and has a relatively compact network structure,but also improves the generalization ability of the network.