Training agents for unknown logistics problems

A methodology on how to prepare agents to succeed on a priori unknown logistics problems is presented. The training of the agents is and can only be executed using a small number of test problems that are taken out of a broad class of generalized logistics problems. The developed agents are then evaluated on unknown instances of the problem class. This work has been developed in the context of last year’s AbstractSwarm Multi-Agent Logistics Competition. The most successful algorithms are presented, and additionally, all participating algorithms are discussed with respect to the features of the algorithms that contribute to their success. As a result, we conclude that such a broad variety of a priori unknown logistics problems can be solved efficiently if multiple different good working approaches are used, instead of trying to find one optimal algorithm. For the used test problems this method can undercut, trivial as well as non-trivial implementations, for example, algorithms based on machine learning.