TGSIM: Third Generation Simulation Dataset

• TGSIM is defined in two senses: as high-resolution empirical vehicle trajectories for studying driver behavior and as a simulation benchmark for MEEG with explicit ground truth.
• In transportation, TGSIM provides 0.1 s interval vehicle trajectories from urban and freeway settings to derive longitudinal kinematics without preset behavioral labels.
• In MEEG, TGSIM serves as a realistic simulation framework integrating frequency-domain models, anatomical priors, and structured sparsity for source connectivity evaluation.

“Third Generation Simulation” or “TGSIM” is used in the cited literature in two distinct senses. In transportation research, “Third Generation Simulation Data (TGSIM)” denotes a publicly available, high-resolution vehicle-trajectory resource used to derive longitudinal interaction variables and local traffic context for human-driven vehicles in mixed urban and freeway settings (Elayan et al., 24 Mar 2026). In MEEG methodology, “TGSIM” refers to a third generation realistic simulation and validation benchmark embedded in BC‑VARETA 1.0, designed for synthetic source activity and source connectivity evaluation in the frequency domain (Gonzalez-Moreira et al., 2018). The shared acronym masks a substantive difference: one resource is an empirical trajectory dataset, whereas the other is a simulation framework with explicit ground truth.

1. Terminological scope and disambiguation

The transportation usage appears in “Behavioral Heterogeneity as Quantum-Inspired Representation,” where TGSIM is the empirical trajectory source used to learn evolving latent representations of driver behavior (Elayan et al., 24 Mar 2026). The MEEG usage appears in the BC‑VARETA 1.0 benchmark description, where the benchmark is described as a coherent, anatomically grounded, frequency-domain, third-generation simulation and evaluation framework for MEEG source activity and connectivity methods (Gonzalez-Moreira et al., 2018).

This distinction matters because the two resources differ in ontology, supervision structure, and evaluation logic. The vehicle-trajectory TGSIM does not provide maneuver labels, driver profiles, or event annotations as ground truth; behavioral profiles are instead learned from trajectories. By contrast, the MEEG benchmark encodes ground truth explicitly through simulated source precision matrices, sparsity masks, and source covariances. A common misconception is therefore to treat “TGSIM” as a single standardized dataset family. The cited papers support a narrower conclusion: the acronym is overloaded across domains, and any technical discussion must specify which TGSIM is intended.

2. Empirical vehicle-trajectory TGSIM in transportation research

In the transportation setting, the dataset comprises two deployments: “Third Generation Simulation Data (TGSIM) Foggy Bottom Trajectories,” available through the U.S. DOT ITS DataHub with DOI 10.21949/1404230, and “Third Generation Simulation Data (TGSIM) I-395 Trajectories,” also available through the ITS DataHub with DOI 10.21949/1404222 (Elayan et al., 24 Mar 2026). The first site is an urban intersection at Foggy Bottom; the second is a freeway segment on I‑395. The paper uses both to capture heterogeneity across environments.

The sampling rate is reported as “0.1 s intervals.” The paper treats TGSIM as providing sufficiently precise, time-aligned trajectories to derive longitudinal kinematics and interactions. It states that the dataset supplies time-stamped positions and speeds sufficient to compute relative speed to a leader, acceleration, and headway, and to derive local traffic context, including density, average speed, and proximity to pedestrians and stop signs. Sensor modalities, coordinate systems, and collection methodology are not specified in that paper.

After preprocessing, the final dataset used for modeling comprises 4,360 human-driven vehicle trajectories and 3,200,397 observations. Of these trajectories, 1,277 come from Foggy Bottom and 3,083 from I‑395. The analysis is restricted to human-driven vehicles; autonomous vehicles are not included. Class distributions, driver demographics, vehicle types, lane metadata, and temporal span are not reported. The paper further notes that TGSIM is presented as newer-generation trajectory data, but it does not explicitly compare the resource to prior datasets such as NGSIM or detail how it differs.

TGSIM as used in this work does not provide maneuver labels, driver profiles, or event annotations as ground truth. This absence of supervisory annotation is central to the modeling choice: the learned behavioral structure is extracted from trajectories rather than matched to predefined taxonomies.

3. Feature derivation and quantum-inspired representation on vehicle TGSIM

The preprocessing pipeline is minimal but specific. The authors state that “Raw trajectories were Gaussian-smoothed and filtered to retain only continuous trajectories” (Elayan et al., 24 Mar 2026). From these trajectories they derive a behavioral vector

$$x_{it} = \begin{bmatrix} \Delta v_{it} & a_{it} & h_{it} \end{bmatrix}^{\!\top}, \quad d = 3,$$

where the components are relative speed to leader, signed acceleration, and headway distance to leader. The context vector has dimension $q=4$ and consists of proximity to pedestrians, proximity to stop signs/signals, local traffic density, and average speed in front and lateral perception zones. The paper defines density and average speed conceptually through perception regions, but does not specify exact radii or perceptual geometry.

Behavioral observations are embedded with nonlinear Random Fourier Features:

$$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$

with $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$ and $b_j \sim \mathrm{Uniform}(0, 2\pi)$, sampled once and fixed across drivers and time steps. The mapped vector is then normalized as

$$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$

Population-level behavioral profiles are represented as density matrices $\rho_k \in \mathbb{R}^{D \times D}$ subject to

$$\rho_k = \rho_k^\top,\quad \rho_k \succeq 0,\quad \mathrm{Tr}(\rho_k) = 1.$$

The profiles therefore satisfy symmetry, positive semidefiniteness, and trace normalization. Context-dependent activation is modeled through a softmax,

$$\pi_k(c_{it}) = \frac{\exp(\beta_k^\top c_{it})} {\sum_{j=1}^K \exp(\beta_j^\top c_{it})},$$

and state evolution combines temporal persistence with a context-weighted mixture. Observation likelihood is quadratic in the normalized RFF embedding, and the sequential update blends the predicted state with the outer product $\tilde{\phi}(x_{it})\tilde{\phi}(x_{it})^\top$. The estimation objective is

$q=4$0

Experimentally, the model is trained jointly on the Foggy Bottom and I‑395 deployments. The reported settings are $q=4$1 and profile counts $q=4$2, with $q=4$3 selected because the mean negative log-likelihood per observation improved from 0.658 at $q=4$4 to 0.629 at $q=4$5, with no further improvement at $q=4$6 (0.629). Optimization uses automatic differentiation in PyTorch. The paper reports a parameter count

$q=4$7

for $q=4$8, $q=4$9, and $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$0, and notes that per-epoch training was reduced to under 15 minutes from approximately 3–4 hours compared to finite differences. Train, validation, and test splits are not specified, and no baseline models are reported.

4. Learned behavioral structure and limitations of vehicle TGSIM analyses

The learned profile spectra exhibit a pronounced low-rank structure (Elayan et al., 24 Mar 2026). Three of the four profiles are effectively rank-1. Profile 3 is multi-modal, with leading eigenvalues approximately 0.7151, 0.2783, 0.0062, and 0.0003. The paper reports the spectrum table as “1: 0.9999, 0.0000, …; 2: 0.9999, 0.0001, …; 3: 0.7151, 0.2783, 0.0062, 0.0003; 4: 0.9999, 0.0001, …”. This suggests that most learned regimes are nearly pure states in the RFF space, while one regime retains substantial internal heterogeneity.

The context activation coefficients $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$1 are interpreted qualitatively in the paper. Profile 2 has strong positive sensitivity to density, with $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$2, and also responds positively to pedestrian and stop proximity; it is activated in dense, interaction-rich urban conditions. Profile 1 shows negative loadings on spatial context, with strongest suppression by density at $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$3, and is activated in low-density, unconstrained conditions. Profile 4 has moderate positive loadings on all context variables and is aligned with steady-state freeway car-following at higher speeds. Profile 3 is relatively context-insensitive, with negative coefficients on density ($$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$4) and average speed ($$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$5), and is interpreted as capturing transitional or heterogeneous states.

Projected eigenmodes support a behavioral characterization. Profile 1 corresponds to a free-flow or low-constraint regime with the largest headways, reported as mean $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$6 m, and near-neutral acceleration. Profile 2 corresponds to a dense-interaction regime with shorter headways, mean $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$7 m, minimal acceleration, lower average speed, and activation near pedestrians, stop controls, and high density. Profile 3 is multi-modal: one mode shows moderate positive acceleration with comfortable spacing, mean $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$8 m, while another shows more assertive following with higher relative speeds and shorter headways, mean $$\phi_j(x) = \cos(w_j^\top x + b_j), \quad j = 1,\dots,D,$$9 m. Profile 4 is described as responsive following, with mean $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$0 m/s, positive acceleration, and moderate headway, consistent with recovering speed behind a slower leader.

The profile geometry is quantified with pairwise Frobenius distances. All off-diagonal distances are greater than zero; Profile 3 is nearest to Profile 1 with $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$1, while the largest separations occur between Profiles 1 and 4 ($w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$2) and between Profiles 1 and 2 ($w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$3). These distances support the claim that the learned profiles are distinct but not equally separated.

Several limitations are explicit or strongly implied. The paper does not detail sensing modalities, coordinate frames, labeling schema, demographic coverage, lane-specific metadata, or raw variable inventories. Noise handling is limited to Gaussian smoothing and continuity filtering; explicit missing-data procedures and measurement-error characterization are not reported. The analysis is also restricted to longitudinal variables and human-driven vehicles, omitting lateral behavior and richer multi-vehicle interaction cues. The authors recommend future additions of lateral and multi-vehicle features, more flexible context mappings, ablations, broader baselines, and transfer testing across sites.

5. TGSIM as a third-generation MEEG simulation benchmark

In BC‑VARETA 1.0, TGSIM denotes a simulation benchmark rather than an empirical observational dataset (Gonzalez-Moreira et al., 2018). Its purpose is to evaluate, under realistic conditions, the joint reconstruction of source activity and source connectivity from MEEG data using third generation methods. The benchmark integrates a frequency-domain generative model, realistic head modeling and lead-field computation, biologically plausible source configurations and connectivity patterns, and quality measures centered on Earth Mover’s Distance and its Cartesian extension.

The model is specified per frequency component $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$4. Let $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$5 be the sensor-space Fourier coefficients and $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$6 the source-space Fourier coefficients for window $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$7. The observation equation is

$w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$8

with lead-field $w_j \sim \mathcal{N}(0, \sigma^{-2} I_d)$9, positive definite sensor correlation matrix $b_j \sim \mathrm{Uniform}(0, 2\pi)$0, and frequency-dependent nuisance variance $b_j \sim \mathrm{Uniform}(0, 2\pi)$1. The source prior is

$b_j \sim \mathrm{Uniform}(0, 2\pi)$2

where $b_j \sim \mathrm{Uniform}(0, 2\pi)$3 is a Hermitian positive definite source precision matrix encoding undirected connectivity. In equivalent notation,

$b_j \sim \mathrm{Uniform}(0, 2\pi)$4

and the sensor cross-spectrum becomes

$b_j \sim \mathrm{Uniform}(0, 2\pi)$5

The benchmark’s “third generation” character rests on full covariance structure at the source level, sparse Hermitian graphical modeling, frequency-domain Bayesian identification of linear dynamical systems, and anatomical priors. BC‑VARETA imposes structured sparsity via a Group LASSO prior directly on $b_j \sim \mathrm{Uniform}(0, 2\pi)$6, with groups defined by anatomical parcels, such as AAL atlas parcellation through FreeSurfer, and with an anatomical probability mask $b_j \sim \mathrm{Uniform}(0, 2\pi)$7 that encodes plausible short-range and long-range connections. Short-range structure is represented through a spatially invariant empirical kernel for connection-strength decay with geodesic distance; long-range structure is informed by probabilistic maps derived from DTI white matter tracts.

The head model is built from an individual subject T1 MRI using FreeSurfer to extract cortex, inner skull, outer skull, and scalp surfaces. Brainstorm then computes the lead-field for a chosen sensor layout through BEM or FEM under the stationary Poisson approximation of Maxwell equations. Example sensor configurations include EEG in the 10‑5 system with $b_j \sim \mathrm{Uniform}(0, 2\pi)$8 sensors, a real EEG example with a 128-channel MEDICID 5 system, and a real MEG example with 248 magnetometers from the Human Connectome Project. Sensor noise is modeled as complex Gaussian with covariance $b_j \sim \mathrm{Uniform}(0, 2\pi)$9, and the benchmark mixes instrumentation noise in sensor space with biological noise in source space projected to sensors.

6. Simulation protocol, validation metrics, and benchmark use

The synthetic data pipeline is explicit and multi-stage (Gonzalez-Moreira et al., 2018). Ground-truth active patches are first chosen on the cortical surface, with no overlap and distance criteria that define short-range and long-range conditions. Short-range requires centroids within the same parcel and centroid–centroid geodesic distance below 5 cm; long-range requires centroids in different parcels and distance above 8 cm. Three patches per trial are used, with variable sizes such as 4, 8, and 24 active generators.

Connectivity mode is then selected among unconnected, randomly-connected, and fully-connected conditions. A Hermitian positive definite source precision matrix $$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$0 is constructed over the active generators by combining a real symmetric part and an imaginary antisymmetric part with Gaussian entries, imposing the sparsity mask $$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$1, and adjusting eigenvalues to ensure positive definiteness. Source realizations are sampled as

$$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$2

and synthetic source covariance is computed from these realizations. Instrumentation noise in sensor space and biological noise in source space are generated, normalized by energy, and scaled to target SNRs of 19 dB, 7 dB, and 0 dB. Synthetic noisy sensor data are then formed by projecting sources through the simulation lead-field and adding both noise components.

A critical design choice is avoidance of inverse crime: simulation and estimation use distinct lead-fields built from two different T1 MRIs, denoted “T1(sim)” and “T1(est)”. Dataset composition is reported in terms of conditions and trials rather than participants. For each condition, 100 random trials of patch centroids are generated. The synthetic empirical data covariance is computed from $$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$3 windows or samples per frequency component. The framework is band-agnostic, although the paper presents alpha-band examples around 10.5–10.6 Hz.

The benchmark emphasizes quantitative evaluation through EMD in source space and Cartesian EMD in connectivity space. For source activity, EMD compares normalized diagonals of ground-truth and estimated source covariances using a transport problem over cortical geodesic distances. For connectivity, CEMD treats connectivity as a distribution over ordered node pairs and uses the Cartesian ground distance

$$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$4

Additional measures include ROC-derived metrics—TPR, TNR, AUC, and F1—for source localization, and Box’s M statistics for covariance or precision similarity. Partial coherence is derived from the precision via

$$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$5

with $$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$6 defined from the diagonal magnitudes of $$\tilde{\phi}(x_{it}) = \frac{\phi(x_{it})}{\|\phi(x_{it})\|}.$$7.

The benchmark is implemented in MATLAB within the BC‑VARETA 1.0 repository at https://github.com/egmoreira/BC-VARETA-toolbox. The paper states that it is “freely available in matlab format” but does not specify a separate dataset DOI or a license. Baseline comparators are eLORETA and the LCMV beamformer. Across simulated short-range and long-range scenarios and across unconnected, randomly connected, and fully connected modes, BC‑VARETA is reported to outperform eLORETA and LCMV in EMD, CEMD, ROC metrics, and Box’s M statistics. A plausible implication is that, in this context, “TGSIM” is more accurately understood as a benchmarked simulation protocol for method validation than as a static dataset in the conventional archival sense.

7. Access, reproducibility, and interpretive cautions

The transportation TGSIM deployments used in driver-behavior modeling are publicly available through the U.S. DOT ITS DataHub, specifically the Foggy Bottom dataset with DOI 10.21949/1404230 and the I‑395 dataset with DOI 10.21949/1404222 (Elayan et al., 24 Mar 2026). The associated codebase for the density-matrix analysis is linked at https://github.com/wissamkontar/Behavioral-Heterogeneity-as-Quantum-Inspired-Representation. Licensing terms, file formats, and additional documentation are not described in that paper, and users are directed to the ITS DataHub entries for those details.

The BC‑VARETA simulation benchmark is available through its MATLAB repository at https://github.com/egmoreira/BC-VARETA-toolbox (Gonzalez-Moreira et al., 2018). The paper indicates the presence of scripts, pseudocode, and functions for reproducibility, but random seed control is not explicitly documented. The benchmark supports both EEG and MEG layouts and can be adapted to different head models and parcellations.

For interpretation, the most important caution is terminological. In transportation, TGSIM denotes empirical vehicle trajectories sampled at 0.1 s and used without native behavioral labels. In MEEG, TGSIM denotes a synthetic, anatomically grounded benchmark with explicit ground truth. The two resources share only an acronym and a broad association with “third generation” methodology. Any citation, comparison, or methodological transfer involving TGSIM therefore requires immediate specification of domain, data-generating process, and ground-truth status.