Speaker：何启明教授，滑铁卢大学

Inviter：姚大成研究员

Title：Generating Samples for Training Neural Networks for GI/G/K Type Queueing Models

Language：English

Time & Venue：2024.09.25 09:30-11:00 思源楼625

Abstract：We introduce a program for generating samples (data) for training neural networks (aka: big models) for GI/G/K type queueing models, which have few analytic solutions. The program is based on matrix-analytic methods. By putting together the CSFP (count-server-for-phase) method, the matrix-geometric solutions for quasi birth-and-death processes, and the matrix-exponential solutions for Markov modulated fluid flow (MMFF) processes, we are able to generate a large number of samples (data), which are dense in the set of all queueing models of interest, for training and validating neural networks for the GI/G/K type queueing models. The effectiveness of the program is compared to existing asymptotic methods. The work sheds lights on the training of neural networks for complex stochastic systems. (Joint work with Haoran Wu, Zhenggao Wu, Haokun Zhao, Fangfang Lyu, and Kiefer Joe Burgess)