Online Linear Optimization for Job Scheduling Under Precedence Constraints

We consider an online job scheduling problem on a single machine with precedence constraints under uncertainty. In this problem, for each trial t = 1, . . . , T, the player chooses a total order (permutation) of n fixed jobs satisfying some prefixed precedence constraints. Then, the adversary determines the processing time for each job, 9 and the player incurs as loss the sum of the processing time and the waiting time. The goal of the player is to perform as well as the best fixed total order of jobs in hindsight. We formulate the problem as an online linear optimization problem over the permutahedron (the convex hull of permutation vectors) with specific linear constraints, in which the underlying decision space is written with exponentially many linear constraints. We propose a polynomial time online linear optimization algorithm; it predicts almost as well as the state-of-the-art offline approximation algorithms do in hindsight. In preliminary experiments, our algorithm runs co...

• Type: Journal
• Year: 2015
• Where: ALT
• Authors: Takahiro Fujita, Kohei Hatano, Shuji Kijima, Eiji Takimoto
• Keywords: ALT 2015, Machine Learning