MAFE: Multi-Agent Fair Environments for Decision-Making Systems

TL;DR

This work formalizes Multi-Agent Fair Environments (MAFE) as a Fair Dec-POMDP defined by an eight-tuple $(\mathcal{N},\mathcal{S},\{\mathcal{A}_n\},\{\mathcal{O}_n\},\mathcal{T}, \gamma, \{c_n^{(R)}\},\{c_n^{(F)}\})$ to integrate Reward Component Functions $\{c_n^{(R)}\}$ and Fairness Component Functions $\{c_n^{(F)}\}$ for dynamic, multi-agent fairness assessment. It introduces three benchmark MAFEs—Loan MAFE, Healthcare MAFE, and Education MAFE—to simulate loan pipelines, health systems, and education-to-employment pathways, respectively, with heterogeneous agents and domain-specific observations, actions, and fairness metrics. A cooperative use case is presented, alongside the Fair Multi-Agent Cross Entropy Method (F-MACEM) to optimize a multi-objective objective combining direct rewards, rate-based terms, and fairness penalties, and to analyze the impact of interventions on disparities over time. The results demonstrate that MAFE-based testbeds enable rigorous evaluation of fairness interventions and coordination strategies, providing practical, domain-spanning platforms for FairAI research and development in socio-technical decision systems.

Abstract

Fairness constraints applied to machine learning (ML) models in static contexts have been shown to potentially produce adverse outcomes among demographic groups over time. To address this issue, emerging research focuses on creating fair solutions that persist over time. While many approaches treat this as a single-agent decision-making problem, real-world systems often consist of multiple interacting entities that influence outcomes. Explicitly modeling these entities as agents enables more flexible analysis of their interventions and the effects they have on a system's underlying dynamics. A significant challenge in conducting research on multi-agent systems is the lack of realistic environments that leverage the limited real-world data available for analysis. To address this gap, we introduce the concept of a Multi-Agent Fair Environment (MAFE) and present and analyze three MAFEs that model distinct social systems. Experimental results demonstrate the utility of our MAFEs as testbeds for developing multi-agent fair algorithms.

Figures (14)

• Figure 1: Illustration of our MAFE definition. (a) A diagram capturing the elements of our MAFE. The actions produce by the model(s) are imported to the environment to be taken by environmental agents. This leads to state transition within the environment that produces a set collection of observations, rewards, and fairness components for each agent which are output by the environment for the model(s) to use to produce actions in the next time step. (b) An example illustrating this process for a healthcare MAFE that particularly captures how the component functions can be used to construct measures of rewards and fairness.
• Figure 2: Distribution plots that illustrate disparities in (a) the qualification score distributions of customers in the Loan Environment, (b) health risk score distributions among geographic sub-populations the Healthcare Environment, and (c) GPA score distributions of students in the Education Environment at the beginning of an episode.
• Figure 3: Plots illustrating the impact of various interventions in each environment, isolating their effects while holding other factors constant. (a) In the Loan MAFE, the effect of 20% debt relief on qualification scores for the full population. (b)-(d) In the Healthcare MAFE, the effects of providing hospital beds, universal health insurance, and unlimited public health investment on mortality rates. (e)-(g) In the Education MAFE, the effects of unlimited tertiary investment, full scholarships, and mentorship on graduation rates for the full population (e) and (f) and the disadvantaged population (g). (h) In the Education MAFE, the effect of unlimited diversity incentives for the Employer Agent on the average utility of workers from disadvantaged groups.
• Figure 4: Learning curves showing realized rewards obtained during training for models with different combinations of reward terms explicitly included in the F-MACEM’s objective function: "Direct"; "Direct + Fair"; or "Direct+Fair+Rate" in the objective.
• Figure 5: Average actions taken by agents over training epochs in Education MAFE.