Working Vacation Scheduling of MX/M/1/N System using Neural Network

Abstract

Optimal scheduling of working vacation (WV) in queueing system is complex because of the stochasticity of arrival-service process. In this study, we suggest a neural network based scheduling (NNS) model because the neural network has the capability of handling the complexity of the system. We focus on the MX/M/1/N system where X is a random variable and the vacation time is dependent on the operation time. The neural network is used to estimate the performance of each action, and the scheduling rule is made from this estimation. Since the target of neural network to train is not obtainable, we contrived a mathematical model to feed the neural network with target. The experimental results show that estimated value from neural network follows the trend of true value, and the performance of NNS was shown to outperform a common exhaustive vacation (EV) scheduling baseline in most cases. We also identified the settings in which we can expect our NNS to achieve high performance.