Agent‑level emulators cut traffic‑injury simulation time by almost 90%
arXiv: 2611.07995v1
Researchers have introduced a new way to speed up detailed population simulations. The idea is to use a trained generative model that can decide the outcome of an event – such as a road‑traffic injury – in one step, instead of running every intermediate event that usually follows the injury. This replaces the long chain of stochastic events with a single, fast sample from the model.
In their test, the team used a conditional tabular generative adversarial network (CTGAN) to replace the road‑traffic‑injury module inside a complex multi‑disease simulation called Thanzi La Onse. The new approach shaved 86.6 % off the computational time needed to run the simulation (95 % bootstrapped confidence interval [85.2 %, 88.1 %]). At the same time, the overall yearly disability‑adjusted life years (DALYs) derived from traffic injuries were recovered with only a 13 % difference compared with the original, full model.
Table listing CTGAN hyper-parameters such as embedding_dim, generator_dim, discriminator_dim, learning rates, batch size, and training epochs.
SDV Parameter
Value
Meaning
embedding_dim
64
Size of categorical embeddings
generator_dim
(256, 256)
Generator hidden layers
discriminator_dim
(256, 256)
Discriminator hidden layers
generator_lr
2e-4
Generator learning rate
generator_decay
1e-06
Generator weight decay
discriminator_lr
2e-4
Discriminator learning rate
discriminator_decay
1e-06
Discriminator weight decay
batch_size
500
Samples per batch
discriminator_steps
1
Discriminator updates per step
log_frequency
True
Log training progress
epochs
500
Training epochs
pac
10
Pack size for training
Table 2. Table 2 lists the CTGAN hyper‑parameters used to replace the detailed road‑traffic‑injury module in the Thanzi La Onse simulation.Molaro, Margherita, Graham, Matthew M., Watson, Oliver, Murray-Watson, Rachel E., Chalkley, Martin, Collins, Joseph H. et al. · arXiv:2611.07995v1 · Table 2: · CC BY 4.0
The authors stress that the method works only when the sequence of intermediate events that follows an injury is detachable – that is, the events affect only the injured individual and not others or the environment. If an injury can influence other agents or create environmental changes before settling, the emulator would not preserve the true dynamics.
Figure 1. Figure 1 illustrates that intermediate event sequences between a road‑traffic injury and its resolution can be detached from the simulation run‑time. The diagram shows how the outcome of the episode can be sampled directly from a trained generator Gθ, eliminating the need to schedule intermediate events that only affect the injured individual.Molaro, Margherita, Graham, Matthew M., Watson, Oliver, Murray-Watson, Rachel E., Chalkley, Martin, Collins, Joseph H. et al. · arXiv:2611.07995v1 · Figure 1: · CC BY 4.0
Because the emulated module retains the full agent‑based framework, it still captures interactions between diseases in the population, while dramatically improving simulation speed.
Agent‑level emulators cut traffic‑injury simulation time by almost 90% | arXiv News
