PSU CS410/510SS: Search and Scheduling
Lecture 7
Case Study: Squeaky-Wheel Optimization

Administrative

• Project Plan
• Project Presentation

Lecture 7

• History
  • Nemhauser et. al. at Georgia State: Lucent cable-assembly problem
  • MILP works somewhat, but needs good ``columns''
  • CIRL agrees to do column generation
  • David Joslin starts thinking about doubleback, comes up with SWO idea
  • SWO turns out to be better by itself
  • MILP OK "tune-up" method for SWO
  • SWO applied to other problems
    • Graph-coloring benchmarks
    • Another shop problem
• Cable assembly problem characteristics
  • Multi-machine shop problem
  • Setups, forbidden setups, release times, deadlines
  • 297 tasks, 13 lines
• Squeaky-Wheel Optimization
  • Basic idea: ``subsearch''/``metasearch''. C.f.
    • Genetic algorithms for tuning local search (e.g. Optiflex)
    • Subsearch in heuristics: (e.g. Parkes ``subsearch'' SAT solver)
  • Greedy heuristic is weak: could ``strengthen''
  • Idea: start with MCF variable ordering, but order ``trouble'' variables earlier
  • Idea: avoid wild oscillations; ``sticky sort'' on blame
  • Blame computation must be ``balanced'' properly
  • Consequences of misassigning blame: effective cycles
• SWO and cable problem
  • Finds good solutions very rapidly
  • Finds best-known solutions quickly
  • Scalability is unclear
• SWO and bin packing
  • Largest-first barely works
  • What might SWO do?