On the optimal design of radial basis function neural networks for the analysis of nonlinear stochastic systems

In this paper, an iterative selection strategy of Gaussian neurons for radial basis function neural networks (RBFNN) is proposed when the RBFNN method is applied to obtain the steady-state solution of the Fokker- Planck-Kolmogorov (FPK) equation. A performance index is introduced to rank neurons. Top rank neurons are selected, leading to a RBFNN with optimal number and locations of Gaussian neurons for the FPK equation under consideration. The statistical properties of the performance index are studied. It is found that the index assigned to the jth neuron is proportional to the probability of the system falling into the small neighborhood of the mean of this neuron as well as proportional to the weight of the neuron. The RBFNN method with the optimally selected neurons is then applied to several challenging examples of nonlinear stochastic systems in 2, 3 and 4 dimensional state space. The RBFNN solutions are also compared with the results of extensive Monte Carlo simulations. It is observed that the RBFNN method with optimally selected neurons by the proposed iterative algorithm is much more efficient than the RBFNN method with uniformly distributed neurons, and is very accurate in terms of the root mean squared (RMS) errors of the FPK equation or the RMS errors of the PDF solution compared with simulation results.