Exponential penalty function control of loss networks (Q1769408)

The so-called ``loss network'' stochastic system consists of a network of resources (facilities), each with a known fixed capacity. Requests for using this network belong to a diverse set of request classes, differing in the arrival rate, the service duration, the resource requirements and the willingness to pay. There is no waiting room (queue), therefore an arriving request must be either admitted into the system for service and assigned an appropriate resource allocation or rejected (lost) at the instant it arrives. An admitted request occupies the allocated resources for the service duration and releases all the resources simultaneously. The objective of the system controller is to design an admission control policy that optimizes an appropriate performance measure of the revenue generated. In this paper the stochastic optimization problem is replaced by a suitably constructed linear program (LP). The optimal solution of this LP yields a target point that is translated into an admission control policy using an exponential penalty function. It is established that this policy is approximately optimal in the limit where individual resource requests are small compared to the total capacity. Moreover, it is shown that this penalty policy performs well in the transient period as well. Thus, in this paper (i) explicit upper bounds for the maximum achievable revenue rate for any time \(t\geq0\) are developed; (ii) an exponential penalty-based admission control policy that is provably approximately optimal for all times \(t\geq0\) in the Halfin-Whitt limiting regime is constructed; (iii) it is shown that the proposed approach can be extended to track arbitrary polyhedral target sets; (iv) results from preliminary simulation studies are included.