Policy Optimization by Neural Network and Its Application to Queueing Allocation Problem
The problem of allocating an arriving customer to one of parallel servers has been actively studied in queueing theory for load balancing in computer networks or in multi-processor systems. To theoretically derive the optimal allocation policy, the assumption of identical servers is usually required. However this assumption is unrealistic in many applications. This paper considers the queueing allocation problem with non-identical servers and multi-class customers. The goal is to optimize the allocation policy with respect to the mean delay of an arbitrary customer. To this end, we represent the allocation policy by a neural network; namely, we allocate an arriving customer according to the output of the neural network, where the inputs to the neural-net are the numbers of queueing customers at each server and the class of the arrival. By using the simulated annealing method, we search the optimal allocation policy in the weight space. Numerical results show that the present procedure significantly reduces the mean delay in comparison to an empirical policy.
KeywordsService Time Allocation Policy Service Time Distribution Simulated Annealing Method Identical Server
Unable to display preview. Download preview PDF.
- M. Ohnishi, “On two control problems of queues,” Journal of the Operations Research Society of Japan, no. 4, pp. 179–184, 1991 (in Japanese).Google Scholar
- J. Takinami, Y. Matsumoto, N. Okino, “Performance evaluation of neural networks applied to queueing allocation problem,” Artificial Neural Nets and Genetic Algorithms, R.F. Albrecht et. al. (eds.), Springer-Verlag, pp. 316–323, 1993.Google Scholar
- S. Markon, H. Kita and Y. Nishikawa, “Reinforcement learning for stochastic system control by using a feature extraction with BP neural networks,” Tech. Rep. of IEICE, NC91-126, pp. 209–214, 1991.Google Scholar