Sequential design with applications to the trim-loss problem

Recent theoretical extensions of a nonparametric sequential design theory first explored by H. Robbins have made this methodology seemingly attractive to a host of decision and learning problems. This work explores these developments in the context of on-line heuristic rule selection for a trim-loss problem which the authors had previously explored in a more conventional setting. The problem is somewhat delicate, being infinite horizon and nonparametric. In particular, as explained herein, it is outside the domain of conventional Bayesian theory as well as currently-popular learning methodology such as neural networks.