PhD Dissertation Proposal Defense: Cen Wang, Deep Learning for Discrete Event Simulations
Content
Speaker
Description
Discrete event simulation is a technique for modeling and analyzing complex stochastic systems, especially in engineering, logistics, and healthcare. Despite recent advances in deep learning, its adoption in discrete event simulation has been slow. This thesis explores how deep learning can solve existing problems and uncover new applications for simulation.
We first introduce Neural Input Modeling (NIM), a generative-neural-network framework that automates simulation input modeling. NIM addresses the challenge of fitting stochastic input-process models to data, especially for non-experts. Its core architecture, NIM-VL, uses a variational-autoencoder (VAE) with Long Short-Term Memory (LSTM) components to learn and reproduce input data distributions while capturing temporal dependencies. NIM can model multivariate processes, handle nonstationary sequences, and perform conditional simulations through conditional NIM (CNIM). Experimental results show that NIM and CNIM effectively automate input modeling, reducing barriers for non-experts and enabling more accurate simulations.
We then develop Graphical Metamodels (GMMs), a new class of simulation metamodels that approximate complex simulation behavior via simple mathematical functions for faster analysis and optimization. Our approach incorporates graph neural networks (GrNNs) to account for the graph structure of simulation models, treating it as an input parameter alongside traditional numerical inputs. We introduce generative graph metamodels (GGMMs) that combine GrNNs with generative neural networks to produce a range of summary statistics and sequences of samples that mimic dynamic outputs. This innovation supports flexible simulation-based prediction and optimization under uncertainty, especially in settings needing quick results. We also propose HiLo, a novel metamodel-training method that reduces computational costs without compromising accuracy. Theoretical and empirical results demonstrate that these methods can accurately model a wide range of simulation scenarios, paving the way for more efficient and flexible simulation studies.
Finally, we propose approaches to simulation optimization problems involving GMMs. GMMs lead to hybrid optimization problems where both continuous numerical parameters, like work rates, and the graph structure, such as warehouse layouts, need to be jointly optimized. We consider an online optimization setting, where, after training a GrNN, we optimize numerical and structural inputs to identify optimal configurations in real-time. We first develop a scalable heuristic optimization method based on Monte Carlo Tree Search. Next, for certain GrNN architectures with ReLU nonlinearity, we formulate this problem as a mixed-integer linear program (MILP) to obtain exact solutions. However, standard MILP solvers face challenges even for small cases. To address this, we will develop a customized solver that can handle the specialized MILP structure. Finally, we will extend traditional offline optimization to hybrid problems using Neural Tangent Kernel theory.
Advisor
Peter J. Haas