Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fairness and Efficiency in Human-Agent Teams: An Iterative Algorithm Design Approach

Mai Lee Chang, Kim Baraka, Greg Trafton, Zach Lalu Vazhekatt, Andrea Lockerd Thomaz

TL;DR

The paper addresses the fairness-efficiency trade-off in human-agent teams by extending fairness metrics to include task preferences from a first-person POV. Through an iterative design that blends large-scale simulations with small human studies, it introduces the fair-equity metric and the Fair-Efficient Algorithm (FEA) to balance individual equity with team productivity. Empirical results show that traditional capability-based fairness can misalign with perceived fairness, especially from a first-person perspective, and that FEA—particularly in mixed Team Types—improves perceived fairness while maintaining competitive efficiency. The work highlights the importance of POV, transparency, and multi-faceted evaluation metrics for designing fair and effective human-agent task allocation systems with practical implications for collaborative AI in real-world settings.

Abstract

When agents interact with people as part of a team, fairness becomes an important factor. Prior work has proposed fairness metrics based on teammates' capabilities for task allocation within human-agent teams. However, most metrics only consider teammate capabilities from a third-person point of view (POV). In this work, we extend these metrics to include task preferences and consider a first-person POV. We leverage an iterative design method consisting of simulation data and human data to design a task allocation algorithm that balances task efficiency and fairness based on both capabilities and preferences. We first show that these metrics may not align with people's perceived fairness from a first-person POV. In light of this result, we propose a new fairness metric, fair-equity, and the Fair-Efficient Algorithm (FEA). Our findings suggest that an agent teammate who balances efficiency and fairness based on equity will be perceived to be fairer and preferred by human teammates in various human-agent team types. We suggest that the perception of fairness may also depend on a person's POV.

Fairness and Efficiency in Human-Agent Teams: An Iterative Algorithm Design Approach

TL;DR

The paper addresses the fairness-efficiency trade-off in human-agent teams by extending fairness metrics to include task preferences from a first-person POV. Through an iterative design that blends large-scale simulations with small human studies, it introduces the fair-equity metric and the Fair-Efficient Algorithm (FEA) to balance individual equity with team productivity. Empirical results show that traditional capability-based fairness can misalign with perceived fairness, especially from a first-person perspective, and that FEA—particularly in mixed Team Types—improves perceived fairness while maintaining competitive efficiency. The work highlights the importance of POV, transparency, and multi-faceted evaluation metrics for designing fair and effective human-agent task allocation systems with practical implications for collaborative AI in real-world settings.

Abstract

When agents interact with people as part of a team, fairness becomes an important factor. Prior work has proposed fairness metrics based on teammates' capabilities for task allocation within human-agent teams. However, most metrics only consider teammate capabilities from a third-person point of view (POV). In this work, we extend these metrics to include task preferences and consider a first-person POV. We leverage an iterative design method consisting of simulation data and human data to design a task allocation algorithm that balances task efficiency and fairness based on both capabilities and preferences. We first show that these metrics may not align with people's perceived fairness from a first-person POV. In light of this result, we propose a new fairness metric, fair-equity, and the Fair-Efficient Algorithm (FEA). Our findings suggest that an agent teammate who balances efficiency and fairness based on equity will be perceived to be fairer and preferred by human teammates in various human-agent team types. We suggest that the perception of fairness may also depend on a person's POV.

Paper Structure

This paper contains 38 sections, 9 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness in Human-Agent Teams
  4. Human-Agent Team Types
  5. Problem Formulation
  6. Fairness Definitions
  7. Reward Function Design
  8. Methodology Overview
  9. Simulation Study Design Overview
  10. User Studies Design Overview
  11. Participants
  12. Design of Agent Teammate
  13. Study Design
  14. Study Procedure
  15. Evaluation of Fair Algorithm
  16. ...and 23 more sections

Figures (5)

  • Figure 1: An example task allocation scenario that falls under our problem formulation. There are different task types (e.g., 0 and 1) and the human teammates may differ in their capabilities and task preferences.
  • Figure 2: User Study #1: An example of the collaborative session showing each teammate's workspace and the two task types used in the user studies: decorate squares cookies (left) and decorate letters cookies (right).
  • Figure 3: Predicted fairness (scaled to rating between 1 and 7) from Simulation Study #1 and participants' ratings from User Study #1: Higher ratings mean more fair. Please see Table\ref{['tab:pilot1_questionnaire1']} for the specific questions. Cap is the abbreviation of capability. Pref is the abbreviation of preference.
  • Figure 4: Simulation study #2 results for the Mixed team type.
  • Figure 5: Predicted fairness (scaled to rating between 1 and 7) from Simulation Study #2 and participants' ratings from User Study #2: Higher ratings mean more fair. Please see Table\ref{['tab:pilot1_questionnaire1']} for the specific questions. Cap is the abbreviation of capability. Pref is the abbreviation of preference.