The Particle Swarm Optimization-radial Basis Function Neural Network Of Artificial Intelligence

This thesis mainly studies the particle swarm optimization algorithm and the radial basis function neural network.As a member of swarm intelligence evolutionary algorithm,the PSO is wildly used in neural network training,function optimization,fuzzy system control and many other fields.And it achieves good results with its simple principle,ease to implement and ability of solving problems with non-derivable node-transfer function or without gradient information,which conventional approaches can't handle.The RBF is a feed-forward neural network which can overcome the heavy dependence on initial value through learning process and avoid converging to local optimum,with its fast computational speed,strong ability of nonlinear mapping and accurate prediction result,thus it has broad application prospects in the reservoir parameter forecast.This paper combines PSO algorithm and RBF neural network to build an artificial intelligence neural network with simple structure,strong stability and high robustness,and applies it to reservoir parameter forecast with seismic attribute of real data.The PSO optimizes problems through collaboration and information sharing among particles of the swarm with advantage of simple principle,ease to implement and few adjusting parameters.However the Standard Particle Swarm Optimization defects in premature convergence,poor ability of local optimization and the like.Against the above,improve the standard algorithm from the following three aspects:updating formula,optimization method and mutation based on cognitive diversity.Introduce linear decreasing inertial weight coefficient,Cauchy distribution random function and Gaussian distribution random function into the updating formula;Adopt asynchronous collaborative optimization strategy with inverted order to improve capacity of solving higher-dimensions problems;Take full advantage of cognitive diversity to make sure the personal-bests mutate at a certain probability to lead the swarm to escape from the local optimum to converge to the global optimum.Benchmark function test indicates the improved PSO is better than other optimization algorithms.Model and real seismic data inversion results show high precision,good lateral continuity and high longitudinal resolution,especially clear depiction among inter-formation deposition,fault and interface,which verify strong global search ability and anti-noise ability as well.First of all,apply Karhunen-Loeve Transform which can reduce dimensionality to preprocessing the input attribute samples.Then,provide network parameter,such as number of hidden node,basis center,variance and the like,by the means of fuzzy clustering analysis.Last,optimize the weights of network using the improved PSO.The UCI data and practical application in prediction of sandstone thickness verify feasibility and practical value of the network with high precision of prediction.