Computational resource management in supervised learning systems

The allocation of computational resources is an important design issue in systems involving the exploration of large search spaces. The authors explore a novel strategy for adjusting adaptively a resource allocation control parameter. This strategy, based on a dynamic envelope model using feedback, modifies the control parameter with a minimum of a priori information. The technique is used in a specific supervised learning system and evaluated by comparing the values generated by the control strategy with optimal control parameter values obtained from extensive measurements. The results demonstrate that the envelope control strategy selects good parameter settings. The strategy requires the specification or adjustment of three parameters: the distance from the envelope vertex to the control parameter, the envelope decay rate, and the angle of the diagonal segment of the envelope