Human-Machine Collaborative Optimization via Apprenticeship Scheduling
Abstract
Scheduling techniques are typically developed for specific industries and applications through extensive interviews with domain experts to codify effective heuristics and solution strategies. As an alternative, we present a technique called Collaborative Optimization via Apprenticeship Scheduling (COVAS), which performs machine learning using human expert demonstration, in conjunction with optimization, to automatically and efficiently produce optimal solutions to challenging real world scheduling problems. COVAS first learns a policy from human scheduling demonstration via apprenticeship learning, then uses this initial solution to provide a tight bound on the value of the optimal solution, there by substantially improving the efficiency of a branch-and bound search for an optimal schedule. We demonstrate this technique on a variant of the weapon-to-target assignment problem, and show that it generates substantially superior solutions to those produced by human domain experts, at a rate up to 10 times faster than an optimization approach that does not incorporate human expert demonstration.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 09, 2016
- Accession Number
- AD1033417
Entities
People
- Julie Shah
- Matthew Gambolay
- Reed Jensen
- Sung-hyun Son
Organizations
- MIT Lincoln Laboratory