This chapter discusses the application of neural networks for the representation of structural responses in earthquake engineering, and their subsequent use in reliability evaluation and optimization for performance-based design. An approach is proposed by means of which the intervening random variables (including the design variables) are separated into two sets: a basic one and, another, grouping all the variables related to the ground motion. Structural responses are deterministically obtained for different combinations of all variables, and neural networks (with the basic set as input) are trained to represent, for example, either the mean or the standard deviation of the responses over the grouped set. Reliability evaluations, and the optimization involved in performance-based design, can then be efficiently performed via simulation. Examples are used to illustrate the approach, and the corresponding advantages are discussed.