Neuro-dynamic programming with recurrent critic for automatic train regulation of metro line

Abstract

Communication based train control bring the possibility of moving block control of metro line. The system's transport capacity and punctuality in service depends on the capability of automatic train regulation. The automatic train regulation problem is essentially nonlinear and stochastic, and limited by operating hours of metro line. Neuro-dynamic programming with recurrent critic is shown able to find near-optimal solution more rapidly and accurately than that with forward critic design. Multilayered perceptrons and back-propagation technique are used to construct the recurrent critic and other components associated with adaptive critic design. Comparisons of automatic train regulation referring to Taipei metro data are made for recurrent critic regulator, forward critic regulator, and linear quadratic regulator. The neuro-dynamic programming with recurrent critic design is shown robust with respect to modeling error and excellent in convergence.