Data-driven robust optimization (Q1702776): Difference between revisions

Problems of robust optimization, which are very popular today, involve constraints including both optimization parameters \(x\) and uncertain parameters \(u\). The last ones must belong to a so-called uncertainty set \(U\), whose choice is crucial for the computational performance of the optimization method being applied. If the distribution \(P\) of \(u\) is unknown one should replace \(U\) by \(U_{\epsilon}\) that implies a probabilistic guarantee for \(P\) at level \(\epsilon\). This \(\epsilon\) describes an allowed small violation in \(P\) and corresponding constraints. The authors suppose to shift accents in constructing \(U_{\epsilon}\) from heuristics to statistical analysis of pre-collected data on \(u\). After advancing and testing of the ypothesis about \(P\) on some significance level \(\alpha\) a convex function is built that describes the set of interest. So the obtained \(U_{\epsilon}\) depends not only on \(\epsilon\), but on \(\alpha\), too. The paper deals with uncertain data with different pre-known properties, and depending on these properties different statistical criteria are used to construct an uncertainty set. The authors examine portfolio management and queueing analysis as examples of the implementation of the proposed techniques. The method has shown a good computational efficiency, as the sets obtained are small enough compared with uncertainty sets obtained by other methods.