Ordinal optimization of \(G/G/1/K\) polling systems with \(k\)-limited service discipline (Q1016394)

The paper deals with a \(G/G/1/K\) polling model consisting of \(Q\) queues \(A_1,\dots,A_Q\) each of them with a finite buffer of size \(K\). The arrival process at the queues is general with a constant rate vector, the service times are i.i.d. The queues are served according to the \(k\)-limited service discipline: the server visits and serves the queues in cyclic order; when visiting queue \(i,\;i = 1, 2, \dots , Q\), the server continues working at this queue until either a predefined number of \(b_i (\leq K)\) customers is served or until the queue becomes empty, whichever occurs first. The switchover time from \(A_q\) to \(A_{q+1}\) is assumed to be runcated normal with mean and variance depending on \(q\). For this model a simulation based ordinal optimization theory based algorithm is proposed to determine the optimal service discipline vector \(b=(b_1,\dots,b_Q)\), \(b_i\leq K\), \(i=1,\dots,Q\), minimizing the sum over all \(Q\) queues of the products of the mean waiting times and waiting costs per time unit and the products of the blocking probabilities and blocking costs. The simulation based proposed algorithm consists of two stages: The first stage is an exploration stage which employs a typical genetic algorithm to search through the optimization set (of all \(k\)-limited discipline vectors) using an off-line trained artificial neural network as a surrogate model for fitness evaluation and selects \(N(=1024)\) roughly good decisions vectors. The second stage is an exploitation stage to find a good enough solution from the \(N\) solutions obtained in Stage 1 with more refined surrogate models. Numerous tests demonstrate the computational properties and quality of the proposed simulation based algorithm. The authors provide a performance analysis of their algorithms.