A game theoretic approach to decentralized multi-project scheduling

In this paper we demonstrate how decentralized multi-project scheduling problems can be solved efficiently by a group of project manager agents playing a simple sequence learning game. In the multi-project scheduling problem, multiple projects, each having a number of activities, must be scheduled. A set of local and global resources are available for carrying out the activities of the projects. It is shown that the sequence learning game improves the best objective function value found (minimal average project delay). In fact, the combination of local reinforcement learning, the sequence learning game and a smart forward-backward implementation of the serial scheduler realizes, on average over all MPSPLIB benchmark instances, a 25% improvement on the best published results.