Understanding Simulation Variation in the Bellevue-Kirkland-Redmond Activity-Based Model
Corresponding Author: Ben Stabler, RSG
Presented By: Ben Sabler, RSG
Unlike traditional aggregate trip-based models, disaggregate activity-based (AB) microsimulation models use Monte Carlo simulation to make choices. Because Monte Carlo simulation relies on random number draws, two runs of the AB model, both with the same inputs but with different random number seeds, will result in different outputs. The purpose of this presentation is to share our findings related to systematically investigating Monte Carlo simulation variation in the new Bellevue-Kirkland-Redmond (BKR) implementation of the PSRC SoundCast AB model.
In order to improve the stability of the model while also decreasing the runtime, we developed an over and undersampling procedure for the AB model. This procedure takes as input a household sample rate for each TAZ and samples households from the synthetic population by income, size, and TAZ to create a new synthetic population. The population sampling procedure was used to expand the population within the BKR region in order to increase the stability of model results while reducing the number of households outside BKR in order to reduce runtimes.
We tested a number of different sampling plans in hopes of finding the best plan for typical applications. We measured variation by summarizing standard deviations and other measure for common travel demand model metrics such as total trips and transit share. These summaries were done at different levels of geography - TAZ, neighborhood, BKR, and PSRC – in order to understand how aggregation impacts the results. Our presentation we will share the results of the many runs, as well as our recommendations for achieving a good balance between model runtime and stability.