Automatic Design of a Hybrid Iterated Local Search for the Multi-Mode Resource-Constrained Multi-Project Scheduling Problem

This paper details our submission to the MISTA 2013 challenge, which deals with the multi-mode resource-constrained multi-project scheduling problem (MRCMPSP). Kolisch and Hartmann [4] recommend metaheuristics as the best-performing methods when tackling the resource-constrained project scheduling problem with a single project and without multiple-modes. Most of these metaheuristics share many similar components, such as neighborhoods used in local search, schedule generation procedures, etc. Given the number of possible combinations of algorithmic components and the various ways in which simple metaheuristics can be combined into a single algorithm, finding the right combination for the MRCMPSP would be, in principle, an arduous task of experimentation by trial-and-error. Instead, we used a recent automatic method [10], which combines (i) a description (given as a grammar) of the space of potentially valid algorithms for a problem and (ii) a method for searching the best algorithm by instantiating algorithms from this grammar. The approach shares some similarities with genetic programming (GP) [11] and grammatical evolution (GE) [3], yet there are crucial differences. First, GP/GE typically attempt to generate programs from very basic components, whereas our approach relies on humans to provide problem-specific components for the particular problem. Second, GP/GE use a tree-based and a codon-based representation, respectively, to instantiate programs from the grammar. Instead, we use a parametric representation, that is, the grammar description is transformed into a number of categorical and numerical parameters, and some of them might be only enabled depending on other parameters (conditional parameters). This transformation requires to specify the maximum number of times that each rule in the grammar can be applied (similar restrictions exist with other representations, given the existence of recursive derivation rules). Finally, GP/GE use evolutionary algorithms to search for the best instantiation of the grammar. Our approach, by contrast, relies on irace [6], an automatic configuration tool typically used for offline parameter tuning. The characteristics of irace makes it ideal to handle complex parameter spaces, with categorical, numerical and conditional parameters. Moreover, irace is designed for handling heterogeneous problem instances.