Abstract
Population synthesis is a fundamental challenge in transport modeling, where realistic representations of micro-agents are required for agent-based simulations. Traditional methods like Iterative Proportional Fitting (IPF) and Bayesian Networks struggle with the "curse of dimensionality" and the problem of "sampling zeros." This paper introduces a methodological framework based on Variational Autoencoders (VAEs) to "grow" pools of micro-agents. By leveraging deep generative modeling, the proposed approach fits high-dimensional joint distributions more efficiently than conventional Gibbs sampling or graphical models. Using a large-scale Danish travel survey, we demonstrate that while traditional methods perform well in low dimensions, the VAE scales significantly better as attribute complexity increases. Furthermore, the VAE generates agents that are statistically similar to the original population yet virtually distinct, effectively addressing the limitations of simple data replication.
The Challenge of Population Synthesis
In urban planning and transport modeling, researchers rely on agent-based models to simulate the daily activities of thousands of synthetic individuals. Creating these populations typically involves expanding a small survey sample (like a travel diary) into a full-scale synthetic city. Historically, this has been achieved through Iterative Proportional Fitting (IPF) or Gibbs sampling. However, these methods face two major bottlenecks:
- The Curse of Dimensionality: As the number of attributes (e.g., age, income, education, car ownership) increases, the number of possible combinations grows exponentially. Traditional graphical models, such as Bayesian Networks, become computationally prohibitive to train in these high-dimensional spaces.
- Sampling Zeros: If a specific demographic combination is missing from the initial survey sample, traditional models often assume it cannot exist. This leads to synthetic populations that are merely repetitive copies of the survey data rather than diverse, realistic representations.
Methodology: Deep Generative Modeling
To overcome these hurdles, this research adopts the Variational Autoencoder (VAE). Unlike standard autoencoders that simply compress and reconstruct data, the VAE regularizes the latent space using Kullback-Leibler divergence. This ensures the model learns a continuous probability distribution rather than discrete data points.
Architecture and Training
The VAE consists of an Encoder ($Q_\theta$) that maps high-dimensional agent data into a low-dimensional latent space, and a Decoder ($P_\phi$) that reconstructs the agent from that space. A key innovation is the reparameterization trick, which allows the model to remain differentiable during training despite the stochastic nature of sampling from the latent distribution.
By handling mixed data types—using cross-entropy loss for categorical variables (one-hot encoded) and mean squared error for numerical variables—the VAE can process complex agent profiles efficiently.
Experimental Evaluation
The model was tested using the Danish National Travel Survey (TU) across three levels of complexity:
- Basic: 4 attributes (27 dimensions).
- Socio: 21 attributes (121 dimensions).
- Extended: 47 attributes (357 dimensions).
The VAE was compared against industry-standard Gibbs Sampling and Bayesian Networks (BN).
Key Results and Insights
- Superior Scalability: While a Bayesian Network took 24 hours to train on 21 variables and failed to converge on 47, the VAE trained in just 10 minutes on a GPU. This makes the VAE the only viable option for high-dimensional population synthesis.
- Addressing Sampling Zeros: The VAE acts as a statistical "smoother." Its bottleneck structure forces it to learn underlying dependencies, allowing it to generate plausible agents that were not present in the original survey. In contrast, Gibbs sampling was found to essentially replicate the training data, failing to provide the diversity needed for robust simulations.
- Statistical Fidelity: In low-dimensional tests, Bayesian Networks slightly outperformed the VAE in fit. However, as dimensionality increased, the VAE maintained high correlation with the true population while traditional methods either crashed or overfitted.
Conclusion and Future Directions
The VAE framework represents a significant step forward in generating richer synthetic populations with more detailed individual characteristics. While the current model occasionally produces logically inconsistent agents (e.g., a child with a PhD), future work will focus on Conditional VAEs (CVAE) for targeted synthesis and the integration of rule-based filters to ensure logical consistency. Additionally, the use of recurrent architectures like LSTMs could extend this framework to generate entire daily activity chains for agents.
Summary updated