We present an approach to verifying the performance of an intelligent control algorithm for which traditional, deterministic verification is not feasible. Our approach is based on statistical learning theory. We develop a classifier based on simulation data to partition the potential operating region of a system under control (here an autonomous helicopter) into acceptable and unacceptable subregions. Statistical learning theory results can then be used to estimate the "generalization" performance of the classifier, providing rigorous bounds on the expected performance. A neural-network-based controller is used to demonstrate the methodology, the outcome of which is an analytical characterization of a "safe" set of maximum velocity and acceleration values under which particular helicopter maneuvers can be reliably executed with acceptable performance.