Skip to content

Welcome to the Fairness-Aware Multi Agent Imitation Learning Project

FAMAIL

FAMAIL is a research project at San Diego State University that addresses spatial inequality in urban taxi services. Using GPS trajectory data from Shenzhen, China, we develop trajectory editing techniques that modify expert driver trajectories to improve fairness metrics — ensuring that all areas of a city receive equitable taxi service — while maintaining the behavioral realism of the original trajectories.

Dr. Xin Zhang (Advisor)
Robert Ashe (Researcher)
San Diego State University · Department of Computer Science

The Challenge

Taxi services in cities like Shenzhen exhibit significant spatial inequality. Certain areas — often wealthier, centrally-located neighborhoods — receive disproportionately more service relative to demand, while other areas — frequently lower-income or peripheral neighborhoods — are systematically underserved. When this inequality correlates with socioeconomic factors, it raises a systemic fairness concern in urban mobility.

Learn more about the problem →

Our Approach

Rather than generating entirely new synthetic trajectories, FAMAIL edits existing expert driver trajectories — applying small, bounded adjustments to real-world GPS data to improve the equity of taxi service distribution. This preserves the expert knowledge embedded in real driver behavior while steering the overall service pattern toward fairness.

1

Identify

Rank trajectories by their contribution to spatial inequality

2

Edit

Modify pickup locations using gradient-based optimization

3

Validate

Verify edited trajectories remain behaviorally realistic

Explore the Project

Research Goals

Four interrelated objectives — from multi-objective optimization to imitation learning.

Research Goals →

Methodology

A two-phase pipeline combining attribution-based trajectory selection with gradient-based editing.

Two-Phase Pipeline →

Objective Function

Three terms — spatial fairness, causal fairness, and trajectory fidelity — in a single differentiable objective.

Objective Function →

Study Area & Data

Real-world GPS taxi data from Shenzhen, China — 50 expert drivers, 4,320 grid cells, 288 daily time buckets.

Study Area →

Fairness

Two complementary lenses: spatial equality of service and causal independence from demographics.

Fairness Definitions →

Discriminator

A Siamese LSTM network that validates whether edited trajectories remain authentic.

ST-SiameseNet →


The FAMAIL Project — 2025–2026