Machine learning by schedule decomposition: prospects for an integration of AI and OR techniques for job shop scheduling

  • Authors:

  • Ulrich Dorndorf;Erwin Pesch;Toàn Phan Huy

  • Affiliations:

  • FB 5 -- Management Information Systems, BWL 3, University of Siegen, 57068 Siegen, Germany;FB 5 -- Management Information Systems, BWL 3, University of Siegen, 57068 Siegen, Germany;FB 5 -- Management Information Systems, BWL 3, University of Siegen, 57068 Siegen, Germany

  • Venue:
  • Advances in evolutionary computing
  • Year:
  • 2003

Quantified Score

Hi-index 0.00

Visualization

Abstract

A survey of recent solution approaches as well as a class of approximation algorithms is provided for solving the minimum makespan problem of job shop scheduling. We briefly review most recent exact approaches as well as neighbourhood search methods and evolution based heuristics. Genetic algorithm-based meta-strategies serve to guide an optimal design of scheduling decision sequences. Simple sequences of dispatching rules for job assignment as well as learning of promising sequences of one machine and multiple job decompositions are considered. Finally, a number of ways for introducing problem-specific knowledge through constraint consistency tests for propagation will be presented. These ideas are applied in a subproblem based constraint propagation approach that learns to find the best bounds for fixing arc directions. Calculation of the initial lower bounds for the subproblems' "best bounds" uses a branch and bound search. Whenever some problem-specific knowledge through constraint propagation leads to a partial solution of the job shop problem, a complete solution can be obtained with either a branch and bound procedure or some heuristic neighbourhood or priority rule based search. Computational experiments show that the approach can find shorter makespans than other local search approaches. The chapter provides an initial framework for a unified solution approach to many combinatorial optimization problems incorporating techniques from artificial intelligence, e.g. evolutionary algorithms and constraint propagation, and operations research, e.g. metaheuristics and branch and bound.