(eng): In this paper a recurrent network, which consists of O(sqrt(m log m)) RBF (radial basis functions) units with maximum norm employing any activation function that has different values in at least two nonnegative points, is constructed so as to implement a given deterministic finite automaton with m states. The underlying simulation proves to be robust with respect to analog noise for a large class of smooth activation functions with a special type of inflexion.