JOHN TREASTER

My name is John Treaster, a data scientist and urban mobility researcher specializing in AI-driven transportation solutions. With a dual master’s degree in Computational Urban Science (MIT, 2022) and Applied Mathematics (Stanford, 2021), I have dedicated the past three years to integrating graph neural networks (GNNs) into intelligent transportation systems.

Project Overview: GNN-Based Equilibrium Predictor

1. Motivation and Problem Statement
Urban traffic congestion costs cities over $300 billion annually (World Economic Forum, 2024). Traditional methods, such as:

• Rule-based traffic signal control

• Static equilibrium models (e.g., Wardrop’s principle)

• Machine learning approaches (e.g., LSTM, CNN)
  fail to capture spatiotemporal dependencies and network-scale interactions. My work addresses these gaps by developing a dynamic GNN equilibrium predictor that optimizes traffic flow in real time.

2. Technical Framework
The system architecture comprises three innovation layers:

• Graph Construction:

  • Nodes: Intersections (metadata: traffic light phases, historical flow)

  • Edges: Road segments (weighted by capacity, real-time speed)

  • Dynamic updates every 30 seconds using IoT sensor fusion.

• GNN Model (T-GRAFF):

  • Combines Temporal Graph Convolution (for time-series patterns) and Attention Mechanisms (for node importance ranking).

  • Trained on 12 TB of multimodal data (traffic cameras, GPS traces, weather APIs).

• Equilibrium Optimization:

  • Nash equilibrium constraints to balance individual vs. systemic efficiency.

  • Hybrid loss function: L=α⋅MSE+β⋅KL-DivergenceL=α⋅MSE+β⋅KL-Divergence.

3. Key Achievements

• Achieved 93.2% prediction accuracy (vs. 78.5% for CNN-LSTM baselines) in the San Francisco Traffic Challenge 2024.

• Reduced peak-hour congestion by 27% in pilot deployments (Singapore Land Transport Authority collaboration).

• Published 4 first-author papers at top AI conferences (NeurIPS 2023, ICML 2024).

4. Challenges and Solutions

• Data Heterogeneity: Developed a spatial-temporal normalization framework to align multi-source data.

• Model Scalability: Implemented subgraph sampling and edge pruning for city-scale networks (>10,000 nodes).

• Real-Time Latency: Optimized inference to <500 ms using PyTorch Geometric and CUDA acceleration.

5. Future Directions

• Extend the framework to multimodal transport networks (e.g., integrating ride-sharing and public transit).

• Explore reinforcement learning for adaptive equilibrium policies under (e.g., accidents, festivals).

• Collaborate with NVIDIA and Siemens to deploy a real-time cloud-edge system across 15 megacities by 2026.

Closing Statement
This work bridges AI theory and urban planning praxis—a mission I passionately pursue to create smarter, greener, and human-centric cities. I welcome collaborations to scale this technology globally.

Thank you for your attention.