PreSimNet

Official code for PreSimNet: A Synergistic Physics-Encoded Deep Learning Framework for Integrated Prediction and Simulation of Car-Following in Mixed-Autonomy Traffic.

PreSimNet models heterogeneous car-following interactions in mixed-autonomy traffic by coupling short-term trajectory prediction with closed-loop physics-guided behavior simulation. The model learns type-guided features for four interaction settings: AV-AV, AV-HV, HV-AV, and HV-HV.

Overview

PreSimNet uses a shared trajectory encoder to extract temporal features and infer the car-following interaction type. The shared representation feeds two complementary heads:

• Open-loop prediction: directly forecasts future following-vehicle positions for short-term trajectory prediction.
• Closed-loop simulation: predicts parameters for physics-based car-following models and iteratively simulates future gap and velocity.
• Type-guided MoE routing: aggregates interaction-specific experts, with Top-2 routing support in the released main model.

Repository Layout

• model/presimnet.py: main PreSimNet model.
• model/model_*_baseline.py: main baseline models used in the paper.
• checkpoints/presimnet_top2/epoch21_e.tar: released PreSimNet Top-2 checkpoint.
• infer_presimnet.py: inference entry point for .npy trajectory windows.
• train_moe_top2.py: PreSimNet training entry point.
• train_baseline.py: baseline training entry point with --model selection.
• evaluate_baseline.py: baseline evaluation entry point.
• analyze_routing_topk.py: Top-2 routing analysis using the original batch-wise evaluation protocol.
• evaluate_position_rmse.py: direct position RMSE evaluation.
• loader2.py: NumPy dataset loader and data schema.
• data/sample/test_data_sample.npy: small example file for format checks.
• docs/figures/article/PreSimNet.png: framework figure used in this README.

Results

The following tables summarize the paper-level metrics.

Trajectory Prediction And Type Classification

Model	Position RMSE (m)	Classification Accuracy (%)	#Params	Inference Time (ms/batch)
Seq2Seq	0.304	97.12	185,413	2.4
Transformer	0.295	98.62	246,085	2.9
CS-LSTM	0.380	95.68	247,525	4.9
STDAN	0.278	97.10	336,165	6.3
BAT	0.272	98.39	417,605	5.1
HLTP	0.280	98.96	527,882	7.4
PreSimNet	0.258	99.23	247,621	2.7

Behavior Simulation

Model	Velocity RMSE (m/s)	Gap RMSE (m)
Personalized IDM	0.359	1.352
PIDL-IDM (Joint)	1.126	1.085
PILSTM-IDM (Joint)	0.606	0.776
PIT-IDM (Joint)	0.855	0.677
PreSimNet	0.342	0.420

Data Format

Input files are NumPy arrays with shape (N, 40, 13). Each sample contains 20 historical steps and 20 prediction steps.

Index	Meaning
0	dataset id
1	trajectory id
2	timestamp
3	car-following interaction type
4	lead vehicle velocity
5	lead vehicle acceleration
6	following vehicle velocity
7	following vehicle acceleration
8	gap
9	headway
10	relative velocity
11	following vehicle position
12	lead vehicle position

Interaction type ids are 0=AV-HV, 1=AV-AV, 2=HV-HV, and 3=HV-AV.

Installation

Install a PyTorch build that matches your CUDA environment, then install the remaining dependencies:

pip install -r requirements.txt

Smoke Test

Run inference with the released checkpoint and sample data:

python infer_presimnet.py \
  --data data/sample/test_data_sample.npy \
  --checkpoint checkpoints/presimnet_top2/epoch21_e.tar \
  --out outputs/sample_predictions.npz

Run a one-epoch training smoke test:

python train_moe_top2.py \
  --train-data data/sample/test_data_sample.npy \
  --epochs 1 \
  --batch-size 16 \
  --max-batches 2 \
  --checkpoint-dir runs/sample/checkpoints \
  --result-dir runs/sample/results

Training

Train PreSimNet:

python train_moe_top2.py \
  --train-data ../data/train_data.npy \
  --epochs 21 \
  --batch-size 512 \
  --num-workers 4 \
  --gamma 0.9

Train a baseline:

python train_baseline.py \
  --model hltp \
  --train-data ../data/train_data.npy \
  --test-data ../data/test_data.npy \
  --epochs 21 \
  --batch-size 512

Available baseline names are seq2seq, transformer, cslstm, stdan, bat, and hltp.

Evaluation

The evaluation scripts default to the paper's original batch-wise RMSE protocol with batch_size=512 and drop_last=True.

Top-2 routing analysis:

python analyze_routing_topk.py \
  --checkpoint checkpoints/presimnet_top2/epoch21_e.tar \
  --data ../data/test_data.npy \
  --out-dir outputs/routing_top2

Direct position RMSE:

python evaluate_position_rmse.py \
  --data ../data/test_data.npy \
  --out outputs/position_rmse.csv

Baseline evaluation:

python evaluate_baseline.py \
  --model hltp \
  --checkpoint checkpoints/hltp_baseline/baseline_epoch21.tar \
  --data ../data/test_data.npy \
  --out outputs/hltp_baseline_test.csv