NIM: GENERATIVE NEURAL NETWORKS FOR MODELING AND GENERATION OF SIMULATION INPUTS

We introduce Neural Input Modeling (NIM), a generative-neural-network framework that exploits modern data-rich environments to automatically capture complex simulation input distributions and then generate samples from them. Experiments show that our prototype architecture NIM-VL, which uses a variational autoencoder with LSTM components, can accurately, and with no prior knowledge, automatically capture a range of stochastic processes, including mixed-ARMA and nonhomogeneous Poisson processes, and can efficiently generate sample paths. Moreover, we show that the outputs from a queueing model with (known) complex inputs are statistically close to outputs from the same queueing model but with the inputs learned via NIM. Known distributional properties such as i.i.d. structure and nonnegativity can be exploited to increase accuracy and speed. NIM has the potential to help overcome one of the key barriers to simulation for non-experts.