Abstract: Based on a fixed sample path, perturbation analysis (PA) offers an estimate for the gradient-d.l(θ)/dθ of performance measure J(θ) with respect to the adjustable parameter θ for stochastic discrete event systems. The PA estimate of dJ(θ)/dθ using fixed length of observation (e. g., L customers) is then put into the stochastic approximation algorithm which recursively optimize the parameter. This is the socalled "Single-Run-Optimization" algorithm. Experiment results show that this kind of algorithms has relatively fast convergence rate. For production systems in series this paper proposes an optimization algorithm which iterates once after every L customers' departure and proves that the algorithm converges to θ which minimizes J(θ).