An agent-based model (ABM) simulates actions and interactions of the synthetic agents to understand the system-level behaviour. The synthetic population, the key input to ABM, mimics the distribution of the individual-level attributes in the actual population. Since individual-level attributes of the entire population are unavailable, small-scale samples are generally used for population synthesis. Synthesizing the population by directly sampling from the small-scale samples ignores the possible attribute combinations that are observed in the actual population but do not exist in the small-scale samples, called ‘sampling zeros’. A deep generative model (DGM) can potentially synthesize the sampling zeros but at the expense of falsely generating the infeasible attribute combinations that should be ‘zero’ in the generated data but exist, called ‘structural zeros’. This study proposes a novel method to ensure that the generation of structural zeros is minimal while recovering the ignored sampling zeros. Two loss functions for regularizing the DGMs are devised to customize the training and applied to a generative adversarial network (GAN) and a variational autoencoder (VAE). The adopted metrics for feasibility and diversity of the synthetic population indicate the capability of generating sampling and structural zeros – lower generation probability of structural zeros and lower generation probability of sampling zeros indicate the higher feasibility and the lower diversity, respectively. Results show that the proposed loss functions achieve considerable performance improvement in the feasibility and diversity of the synthesized population over traditional models. The proposed VAE additionally generated 23.5 % of the population ignored by the sample with 79.2 % precision (i.e., the generation ratio of structural zeros and the total samples is 20.8 %), while the proposed GAN generated 18.3 % of the ignored population with 89.0 % precision. The proposed improvement in DGM generates a more feasible and diverse synthetic population. Since synthesizing the population is the first stage of ABM, the proposed approach improves the overall accuracy of the ABM by circumventing the error propagation to later modelling stages.

基于代理的模型 (ABM) 模拟合成代理的动作和交互以了解系统级行为。合成种群是 ABM 的关键输入，模拟了实际种群中个体水平属性的分布。由于无法获得整个种群的个体水平属性，因此通常使用小规模样本进行种群合成。通过直接从小规模样本中采样来合成总体忽略了在实际总体中观察到但在小规模样本中不存在的可能属性组合，称为“采样零点”。深度生成模型 (DGM) 可以潜在地合成采样零点，但代价是错误地生成不可行的属性组合，这些组合在生成的数据中应该是“零”但存在，称为“结构零点”。本研究提出了一种新方法，以确保在恢复被忽略的采样零点时结构零点的生成最少。设计了两个用于正则化 DGM 的损失函数来定制训练，并将其应用于生成对抗网络 (GAN) 和变分自动编码器 (VAE)。采用的合成种群可行性和多样性指标表明生成采样和结构零点的能力——较低的结构零点生成概率和较低的采样零点生成概率分别表示较高的可行性和较低的多样性。结果表明，与传统模型相比，所提出的损失函数在合成种群的可行性和多样性方面取得了相当大的性能改进。拟议的 VAE 额外生成了 23.5% 的被样本忽略的人口，精度为 79.2%（即，结构零点与总样本的生成比率为 20.8%），而拟议的 GAN 生成了 18.3% 的被忽略的人口，精度为 89.0%精确。在 DGM 中提出的改进产生了更可行和多样化的合成种群。由于合成总体是 ABM 的第一阶段，因此所提出的方法通过避免错误传播到后期建模阶段来提高 ABM 的整体准确性。在 DGM 中提出的改进产生了更可行和多样化的合成种群。由于合成总体是 ABM 的第一阶段，因此所提出的方法通过避免错误传播到后期建模阶段来提高 ABM 的整体准确性。在 DGM 中提出的改进产生了更可行和多样化的合成种群。由于合成总体是 ABM 的第一阶段，因此所提出的方法通过避免错误传播到后期建模阶段来提高 ABM 的整体准确性。