ashley222/standalone-traffic-brain

Standalone Traffic Brain

A compact recurrent Global Workspace model for 15-minute-ahead traffic-flow forecasting across 36 road sensors. It was trained with Gaussian noise and sensor dropout to remain useful when traffic sensors are degraded or missing.

Evaluation

Three fixed-seed runs on the official UCI Traffic Flow Forecasting test split:

Condition	Small MLP MAE	This model MAE
Clean	0.03272	0.03504
Gaussian noise (0.15)	0.03479	0.03531
20% sensors missing	0.06387	0.04811

The model prioritizes fault tolerance. It does not beat the MLP on clean data, but reduces missing-sensor MAE by about 24.7% in the multi-seed evaluation.

Dataset

Trained on UCI Traffic Flow Forecasting (Zhao, 2019), DOI 10.24432/C57897, licensed CC BY 4.0. The data contains traffic readings from 36 sensor locations along highways in Northern Virginia/Washington, D.C.

Limitations

• The model was evaluated on one traffic network and should not be assumed to generalize to other cities without retraining.
• The official test partition was consulted during iterative model development; metrics should be treated as development-benchmark results, not a pristine blind evaluation.
• Persistent context is disabled operationally by resetting between independent batches. Chronological context needs task-specific temporal training.
• This is a research model, not a safety-critical traffic-control system.