摘要：AbstractIn classical paradigm of model identification, a single prediction value is returned as a point estimate of the output. Recently, the interval prediction model (IPM) has been receiving increasing attentions. Different from generic models, an IPM gives an interval of confidence as the prediction that covers the majority of training data while being as tight as possible. However, due to the randomness of sampling training data, the reliability of IPM constructed is uncertain. In this paper, we focus on a general class of IPMs where a fraction of data samples can be discarded to pursue robustness, and establish an appropriate a posteriori reliability guarantee. It relies on counting the “decisive” constraints associated with the optimal solution, and generally leads to reduced conservatism and better estimation performance than the existing performance bounds. Moreover, the guarantee holds irrespective of the data generation mechanism, which informs the decision maker of the prediction confidence in the absence of precise knowledge about data distribution. Its effectiveness is illustrated based on numerical examples.
