Neighborhood synthesis from an ensemble of MIP and CP models

Abstract

In this paper we describe a procedure that automatically synthesizes a neighborhood from an ensemble of Mixed Integer Programming (MIP) and/or Constraint Programming (CP) models. We move on from a recent paper by Adamo et al. (2015) in which a neighborhood structure is automatically designed from a (single) MIP model through a three-step approach: (1) a semantic feature extraction from the MIP model; (2) the derivation of neighborhood design mechanisms based on these features; (3) an automatic configuration phase to find the “proper mix” of such mechanisms taking into account the instance distribution. Here, we extend the previous work in order to generate a suitable neighborhood from an ensemble of MIP and/or CP models of a given combinatorial optimization problem. Computational results show relevant improvements over the approach considering a single model.