What is FAMAIL?

FAMAIL (Fairness-Aware Multi-Agent Imitation Learning) takes a novel approach to the urban taxi fairness problem: rather than generating entirely new synthetic trajectories, we edit existing expert driver trajectories to improve fairness.

Trajectory Editing vs. Generation

This editing approach is central to the project's philosophy:

Approach	Description	FAMAIL?
Trajectory Generation	Create entirely new synthetic trajectories from scratch	✗
Trajectory Editing	Apply small, bounded adjustments to real expert trajectories	✓

Why Editing?

Editing is preferred for several reasons:

1. Preserves expert knowledge. Driving patterns, route choices, and temporal behaviors are retained from the original human drivers.

2. Maximizes efficiency. Not all trajectories contribute equally to global unfairness, so only the most impactful trajectories are modified.

3. Bounded modifications. Spatial constraints limit how far pickup locations can move, ensuring modifications remain realistic.

4. Enables fidelity validation. A neural network discriminator can verify that a modified trajectory still "looks like" it came from the same driver.

The FAMAIL Pipeline

1

Identify Unfair Trajectories

Which trajectories contribute most to spatial inequality?

→

2

Edit Pickup Locations

Move pickup locations toward underserved areas

→

3

Validate Fidelity

Does the edited trajectory still look realistic?

Long-Term Vision

The edited trajectories are ultimately intended to serve as training data for imitation learning agents. By training agents on fairness-improved demonstrations, we aim to produce autonomous taxi policies that naturally distribute service more equitably — embedding fairness directly into learned driving behavior.

Expert GPS Trajectories
        │
        ▼
  FAMAIL Trajectory Editing
  (Identify → Edit → Validate)
        │
        ▼
  Fairness-Improved Trajectories
        │
        ▼
  Train Imitation Learning Agents
  on Fairer Expert Demonstrations

---