Consider a single machine that produces a single item in an M/G/1 environment with backlog. There will be n classes of customers differing from each other only by their backlog costs and a linear (in both time and items) holding cost. A central planner will face two types of decisions: allocation—whether to allocate an item and to whom; andproduction—whether to produce another item or to remain idle. The central planner’s objective is to minimize the average cost per unit of time.
Given the stationary nature of the problem and memorylessness of the arrival process, a simple regeneration argument establishes a stationary base stock level policy is the optimal production policy. In such cases, a make-to-stock strategy is used until the base stock level, S, is reached; production is stopped only when there are S items in stock. Of course, finding the optimal base stock level, S∗, still remains an issue.
Three policies considered in the literature on controlling make-to-stock systems are First Come First Served (FCFS), Multi-level rationing (MR), and strict priority policies. In the FCFS policy, when a decision to allocate an item is made, the items are allocated to the customer who has been waiting the longest independent of his type. The FCFS policy is not optimal because, e.g., when there are backlogs, it is better to allocate items to customers with higher backlog cost. The MR policy is defined by a sequence of rationing levels 0=L1≤L2⋯≤Ln≤Ln+1=S, such that an item is allocated to a class j≥i customer if and only if the stock level is at least Li. Within classes j≥i items are allocated in a FCFS fashion. The strict priority policy is a special case of the MR policy where items are allocated FCFS as long as there is inventory and are prioritized when there is backlog, i.e., 0=L1=L2⋯=Ln≤Ln+1=S.
The problem of minimizing the average cost of a make-to-stock system with priorities was first investigated in [6]. This work and much of the following one focused on theM/M/1 settings. In [10] it shows that the MR policy is optimal for the M/M/1 make-to-stock queues. Recently, [2] provided the exact analysis of the strict priority and MR policies for the M/G/1 settings. They characterize the backlog cost for each customer class by considering an M/G/1 backlog queue for this class. These backlog queues were calibrated to have backlog cost identical to the one in the original system by extending the results of [7] with respect to the distribution of the residual service time observed by arrivals. Recently, [5] provided a more intuitive derivation for these distributions. [2] notes that the MR policy may not be optimal in these settings.
In the M/M/1 settings information on the time spent in a specific inventory and backlog levels have no value due to memorylessness of the arrival and production processes, but such information may be valuable in the M/G/1 settings. For example, consider the case of deterministic service time of length M=5 h. Assume that we observe that an hour after production starts a low priority customer arrives and that the inventory level is such that it is optimal to backlog this customer. If there are no additional high priority arrivals after 3 more hours we know production will be completed within an hour. We can reduce the cost by allocating an item from stock to the low priority customer (even before the production completion). While being able to relay on full information may be optimal, this requires a continuous review of the system and, thus, is hard to implement. We therefore focus on finding the optimal policy within the class of policies that only take actions at arrival and service completion times, based on the information on the inventory and backlog positions at these times. Such policies are more applicable than polices that require full information. Note that policies within this class are static and do not use information on the time passed since the last arrival or since production completion. However, such policies can use information on the inventory and backlog to estimate the time to the next production completion and then use this estimation to reduce costs. In addition, the following analysis and results below can be extended in a straightforward manner to cases where information on the time elapsed since the last production starts is available in a straight forward manner using the distribution of residual service time. Applying the resulting controls would reduce costs in these settings. We comment on this straightforward extension in the conclusion.
We propose an Extended MR (EMR) policy that improves the control of the system by exploiting the information on the residual service time, given the level of backlog. We use queue decomposition of the M/G/1 system with a state dependent arrival rate, in the spirit of [1]. We show that the EMR policy may reduce the costs of an M/G/1 make-to-stock system and it is optimal when production times have an Increasing