Abstract: A two stage RBF network training method for dynamic systems was proposed. Taking its inspiration from dynamic nature and adaptability of natural immune system, in the method, the RBF hidden units are differentiated into two types, and undergo different training stages according to the stable and mutation part of dynamic systems. Through adaptive immune mechanism, the method greatly reduces the computational complexity and improves the dynamic tracing ability. The computer simulations demonstrate that the proposed method strikes a good balance between precision and convergence speed, and achieves very high performance.

Key words:  radial basis function (RBF) network, immune system, adaptive immune mechanism, dynamic system