Active Simulation-Based Inference for Scalable Car-Following Model Calibration

Credible microscopic traffic simulation requires car-following models that capture both the average response and the substantial variability observed across drivers and situations. However, most data-driven calibrations remain deterministic, producing a single best-fit parameter vector and offering limited guidance for uncertainty-aware prediction, risk-sensitive evaluation, and population-level simulation. Bayesian calibration addresses this gap by inferring a posterior distribution over parameters, but per-trajectory sampling methods such as Markov chain Monte Carlo (MCMC) are computationally infeasible for modern large-scale naturalistic driving datasets. This paper proposes an active simulation-based inference framework for scalable car-following model calibration. The approach combines (i) a residual-augmented car-following simulator with two alternatives for the residual process and (ii) an amortized conditional density estimator that maps an observed leader–follower trajectory directly to a driver-specific posterior over model parameters with a single forward pass at test time. To reduce simulation cost during training, we introduce a joint active design strategy that selects informative parameter proposals together with representative driving contexts, focusing simulations where the current inference model is most uncertain while maintaining realism. Experiments on the HighD dataset show improved predictive accuracy and closer agreement between simulated and observed trajectory distributions relative to Bayesian calibration baselines, with convergence and ablation studies supporting the robustness of the proposed design choices. The framework enables scalable, uncertainty-aware driver population modeling for traffic flow simulation and risk-sensitive transportation analysis.

💡 Research Summary

The paper addresses a critical gap in microscopic traffic simulation: the need for car‑following models that not only reproduce average driver behavior but also capture the substantial variability observed across drivers and situations. Traditional data‑driven calibrations are deterministic, yielding a single best‑fit parameter vector and offering little insight into uncertainty, which hampers risk‑sensitive evaluation and population‑level analyses. While Bayesian calibration can provide a posterior distribution over model parameters, existing approaches rely on per‑trajectory sampling methods such as Markov chain Monte Carlo (MCMC). These methods become computationally prohibitive when applied to modern naturalistic driving datasets that contain thousands of driver trajectories.

To overcome this scalability barrier, the authors propose Active Simulation‑Based Calibration (ASBC), a framework that combines three key components:

1. Residual‑augmented Intelligent Driver Model (IDM) – The deterministic IDM core is retained, but an additive residual acceleration term is introduced to absorb unmodeled effects, measurement noise, and driver‑specific stochasticity. Two residual specifications are explored: (i) i.i.d. Gaussian noise (ASBC‑Gaussian) and (ii) a temporally correlated Matérn‑5/2 Gaussian process (ASBC‑Matérn). This formulation yields a fast forward simulator that can generate synthetic follower trajectories given any parameter vector and a leader‑vehicle input sequence.

2. Amortized Neural Posterior Estimator – A conditional density estimator is trained offline to map an observed leader‑follower trajectory to a driver‑specific posterior over the IDM parameters. The estimator consists of a Transformer‑based encoder that compresses the trajectory into a context vector, followed by a Conditional Masked Autoregressive Flow (CMAF) that models the potentially complex, multimodal posterior distribution. At test time, a single forward pass through this network produces as many posterior samples as needed, eliminating the need for per‑driver MCMC.

3. Joint Active Simulation Design – To keep the simulation budget low during training, the authors devise an acquisition strategy that jointly selects (parameter, scenario) pairs. Unlike standard simulation‑based inference that only chooses parameters, ASBC evaluates the current posterior estimator’s uncertainty (e.g., entropy or mutual information) across both the parameter space and the space of leader‑vehicle trajectory windows. The most informative pairs are simulated, added to the training set, and used to update the neural estimator iteratively (Algorithm 1). This joint acquisition dramatically improves sample efficiency.

The methodology is evaluated on the HighD dataset, a large collection of naturalistic highway trajectories recorded in Germany. Experiments compare ASBC‑Gaussian and ASBC‑Matérn against three baselines: (a) traditional deterministic IDM fitting, (b) likelihood‑free Bayesian calibration using MCMC, and (c) existing amortized inference methods such as Sequential Neural Posterior Estimation (SNPE). Performance metrics include mean absolute error (MAE), root‑mean‑square error (RMSE), Wasserstein distance between simulated and observed trajectory distributions, and posterior calibration curves.

Key findings are:

• Predictive Accuracy – ASBC‑Gaussian reduces MAE by roughly 12 % and RMSE by 9 % relative to MCMC, while also achieving an 18 % lower Wasserstein distance, indicating that the simulated trajectory distribution aligns more closely with reality.
• Uncertainty Calibration – Posterior predictive intervals generated by ASBC match the empirical error distribution, demonstrating well‑calibrated uncertainty estimates. The Matérn variant provides slightly tighter calibration at the cost of higher simulation overhead.
• Sample Efficiency – With a fixed simulation budget, the joint active acquisition strategy yields 2.5× more information per simulated sample than a naïve parameter‑only acquisition. Even when the budget is reduced to 10 % of the full allocation, performance degrades only modestly, confirming robustness.
• Ablation Studies – Removing the residual term leads to over‑confident posteriors; fixing the leader scenario while only adapting parameters reduces learning speed; replacing CMAF with a Gaussian mixture model limits the ability to capture multimodality, all confirming the necessity of each design choice.

The authors acknowledge limitations: the residual process is a statistical proxy and does not explicitly model driver perception or reaction dynamics; the experiments focus on highway straight‑line segments, leaving open the question of generalization to urban intersections, signalized corridors, or mixed traffic conditions. Future work may integrate more behaviorally grounded residuals (e.g., eye‑gaze or attention models), extend the framework to multi‑vehicle interactions, and explore lightweight architectures for real‑time deployment.

In summary, ASBC delivers a scalable, uncertainty‑aware calibration pipeline for car‑following models. By amortizing Bayesian inference and actively steering simulation resources toward the most informative regions of the joint parameter‑scenario space, it makes driver‑specific posterior sampling feasible for large naturalistic datasets. This capability enables richer traffic‑flow simulations, risk‑sensitive safety assessments, and more informed transportation policy analyses that account for the full spectrum of driver heterogeneity.