When do agents outperform centralized algorithms?

Multi-agent systems (MAS) literature often assumes decentralized MAS to be especially suited for dynamic and large scale problems. In operational research, however, the prevailing paradigm is the use of centralized algorithms. Present paper empirically evaluates whether a multi-agent system can outperform a centralized algorithm in dynamic and large scale logistics problems. This evaluation is novel in three aspects: (1) to ensure fairness both implementations are subject to the same constraints with respect to hardware resources and software limitations, (2) the implementations are systematically evaluated with varying problem properties, and (3) all code is open source, facilitating reproduction and extension of the experiments. Existing work lacks a systematic evaluation of centralized versus decentralized paradigms due to the absence of a real-time logistics simulator with support for both paradigms and a dataset of problem instances with varying properties. We extended an existing logistics simulator to be able to perform real-time experiments and we use a recent dataset of dynamic pickup-and-delivery problem with time windows instances with varying levels of dynamism, urgency, and scale. The OptaPlanner constraint satisfaction solver is used in a centralized way to compute a global schedule and used as part of a decentralized MAS based on the dynamic contract-net protocol (DynCNET) algorithm. The experiments show that the DynCNET MAS finds solutions with a relatively lower operating cost when a problem has all following three properties: medium to high dynamism, high urgency, and medium to large scale. In these circumstances, the centralized algorithm finds solutions with an average cost of 112.3% of the solutions found by the MAS. However, averaged over all scenario types, the average cost of the centralized algorithm is 94.2%. The results indicate that the MAS performs best on very urgent problems that are medium to large scale.