Multi-objective, multi-attribute fleet sizing in a dynamic and stochastic environment: A data-driven approach

We present a method for sizing a vehicle fleet in operational contexts in which fleet performance is measured along multiple dimensions. One premise of the method is that decision-makers regarding fleet composition are interested in fleets that peform well across multiple seasons and in changing demand markets. Another premise is that daily operations are dynamic and stochastic in that vehicle routing decisions must be determined with incomplete information of future customer requests for service that day. A third premise is the existence of a solver for the daily dynamic planning problem that the method can use in a black-box fashion. Based on these premises, we present a heuristic framework for generating a predictive model of fleet performance. The method involves sampling operational settings across seasons and demand markets and executing a black-box planning tool for different fleet compositions to generate solutions and corresponding performance metric values. These settings and values are used to establish a training data set to which a prediction model is fitted. Once fitted, the decision-maker can use such a prediction model to quickly determine the fleet size and attributes that are likely to perform as desired on the performance measures of interest. To demonstrate the effectiveness of the proposed method we constructed a carefully curated data set from publicly available data sources to simulate the operational context of a grocery home delivery service. In that context, fleet vehicles can have multiple compartments to support the transportation of different food products that require storage at different temperate ranges. Thus, fleet sizing decisions involve both the number of vehicles and the size of each compartment within a vehicle. With an extensive computational study and analysis we illustrate that the proposed heuristic approach produces prediction models, both regression models and neural networks, that exhibit strong predictive power and can effectively inform fleet sizing decisions.