Comparing Methods for Generating a Two-Layered Synthetic Population

Abstract

Synthetic population is used in many transport models ranging from trip-based, hybrid trip, tour-based, and activity-based models. As mobility decisions depend on both individuals’ characteristics and family situation, generating a two-layered population that takes into account not only the individual level but also household level is essential. In the literature, three main categories of methods for two-layered population generation have been proposed. These categories are synthetic reconstruction (SR), combinatorial optimization (CO), and statistical learning (SL). SR and CO methods produce synthetic populations by means of replicating individuals, whereas SL methods generate a population following a joint probability estimation. However, selecting a generation process is not straightforward as it depends on input data and synthetic population characteristics. To the best of our knowledge, no clear methodology for selecting between these methods exists. The main objectives of this paper are to provide (1) a detailed description of the available methods, (2) a comparison between these methods, and (3) a decision-making procedure for selecting between these methods. The description and comparison of the methods relies on different criteria: marginals availability, sample size, number of potential attributes that can be handled, population size to generate, possibility of zero-cell problem, and so forth. The advantages and shortcomings of each method are illustrated, and method performance is assessed. The decision-making procedure is carried out through the proposal of a decision tree. Researchers and practitioners have now access to a comprehensive and unified framework to select the appropriate method depending on available data and features of their modeling purposes.

A synthetic population for agent-based modelling in Canada

“…Many methods have been proposed to generate synthetic populations 28 : Synthetic reconstruction like Iterative Proportional Fitting (IPF) 29 and Iterative Proportional Updating (IPU) 30 , which combines sample data and aggregate local statistics to compute the weights reflecting each sample individual’s representativeness in the local zone. Müller 31 proposed a Hierarchical IPF method which sample the hierarchical PUMF to directly generate a synthetic population of households and individuals.…”

Impact of the Geographic Resolution on Population Synthesis Quality

“…Arentze et al used relation matrices to convert marginal constraints at the person level to additional household-level constraints before using the IPF basic procedure to estimate household joint distributions [14]. However, these methods do not fit households and people distributions simultaneously, and thus do not warrant their consistency [15].…”

“…As the mobility behaviors are determined both by people and households' characteristics [16][17][18], multilevel synthesizers are proposed. The multilevel synthesizers try to fit both households and people distributions by reweighting households according to their compositions of individuals [15]. The multilevel synthesizers can be divided into three categories: synthetic reconstruction, combinatorial optimization, and statistical learning [15,19].…”

“…Bar-Gera et al [21] used entropy optimization to fit households and people distributions simultaneously while minimally altering initial households' weights [22]. Generalized raking [23] can also be used for multilevel fit by distance functions minimization [15]. Fournier et al tried to achieve multilevel fit using optimization-based reweighting approaches, such as non-negative least squares, non-negative least deviation, and cyclical coordinate descent [5].…”

“…For each trial, a goodness-of-fit indicator is calculated [26]. If the fit is improved, the new household is kept; otherwise, it is disregarded [15].…”