Upscaling Agent-Based Discrete-Choice Transportation Models Using Artificial Neural Networks

Agent based models (ABMs) can be used for simulating consumer transportation discrete choice modeling, while incorporating the effects of heterogeneous agent behaviors and social influences. However, the application of ABMs at large-scales may be computationally prohibitive (e.g., for millions of agents). In an attempt to harness the modeling capabilities of ABMs at large scales, we develop a recurrent artificial neural network (ANN) to replicate nonlinear spatio-temporal discrete choice patterns produced by a spatially-explicit ABM with social influence. This particular ABM has been developed to model consumer decision making between purchasing a Prius-like hybrid or plug-in hybrid electric vehicle (PHEV) for a given geographic region (e.g., city or town). Our goal is to see if an ANN trained at the city scale can operate as a “fast function approximator” to estimate nonlinear dynamic response functions (e.g., fleet distribution, environmental attitudes, etc.) based on city-wide attributes (e.g., socio-economic distributions). Recurrent feedback connections were added to the ANN to leverage the temporal history and correlations and improve forecasts in time. Outputs from the city-scale ABM, run for a variety of population sizes and initial and input conditions, were used to train and test the ANN. Initial results suggest the ABM may be replaced by ANNs that interact with each other and other agents (e.g., manufacturing agents) to investigate PHEV penetration at the national scale.