We present tractable algorithms to assess the sensitivity of a stochastic dynamic fleet management model to fleet size and load availability. In particular, we show how to compute the change in the objective function value in response to an additional vehicle or an additional load introduced into the system. The novel aspect of our approach is that it does not require multiple simulations with different values of the model parameters, and in this respect it differs from trial-anderror-based "what-if" analyses. Numerical experiments show that the proposed methods are accurate and computationally attractive.
IntroductionGiven a vehicle fleet and a stochastic process characterizing the load arrivals in a transportation network, the primary objective of fleet management models is to make the vehicle repositioning and vehicle-to-load assignment decisions so that some performance measure (profit, cost, deadhead miles, number of served loads, etc.) is optimized. However, besides making these vehicle repositioning and assignment decisions, an important question that is commonly overlooked by many fleet management models is how the performance measures would change in response to a change in certain model parameters. For example, freight carriers are interested in how much their profits would increase if they introduced an additional vehicle into the system or if they served an additional load on a certain traffic lane. Railroad companies want to estimate the minimum number of railcars that is necessary to cover the random shipper demands. The Airlift Mobility Command is interested in the impact of limited airbase capacities on the delayed shipments. Answering such questions, in one way or another, requires sensitivity analysis of the underlying fleet management model responsible for making the vehicle allocation decisions.In this paper, we develop efficient sensitivity analysis methods for a stochastic fleet management model previously developed in Godfrey and Powell (2002a, b). This model formulates the problem as a dynamic program, decomposing it into time-staged subproblems, and replaces the value functions with specially structured approximations that are obtained through an iterative improvement scheme. Two aspects of this model are crucial to our work: (1) Due to the special structure of the value function approximations, the subproblem that needs to be solved for each time period is a min-cost network flow problem. This enables us to use the well-known relationships between the sensitivity analyses of min-cost network flow problems and min-cost flow augmenting trees. In particular, we can use the fact that the change in the optimal solution of a mincost network flow problem in response to a unit change in the supply of a node or a unit change in the upper bound of an arc is characterized by a min-cost flow augmenting path.(2) Letting be the set of time periods in the planning horizon, just as the value functions V t · t ∈ describe an optimal vehicle allocation policy through the so-called optimality equation (se...