Computational Evaluation of Data Driven Local Search for MIP Decompositions

Driven by the perspective use in decomposition based general purpose solvers, we tackle the issue of improving Dantzig-Wolfe decomposition patterns for generic Mixed Integer Programs (MIP). In particular, we consider the scenario in which a MIP instance and its decomposition are given as input and we address the task of manipulating such decomposition by observing only static algebraic components, with the aim of producing better computational performance features (tighter bounds and comparable computing times). We propose a local search algorithm guided by data driven models and evaluate its performance on MIPLIB instances while starting from decompositions given by either static or data driven detectors.