MAFE: Enabling Equitable Algorithm Design in Multi-Agent Multi-Stage Decision-Making Systems

Algorithmic fairness is often studied in static or single-agent settings, yet many real-world decision-making systems involve multiple interacting entities whose multi-stage actions jointly influence long-term outcomes. Existing fairness methods applied at isolated decision points frequently fail to mitigate disparities that accumulate over time. Although recent work has modeled fairness as a sequential decision-making problem, it typically assumes centralized agents or simplified dynamics, limiting its applicability to complex social systems. We introduce MAFE, a suite of Multi-Agent Fair Environments designed to simulate realistic, modular, and dynamic systems in which fairness emerges from the interplay of multiple agents. We demonstrate MAFEs across three domains – loan processing, healthcare, and higher education – that support heterogeneous agents, configurable interventions, and fairness metrics. The environments are open-source and compatible with standard multi-agent reinforcement learning (MARL) libraries, enabling reproducible evaluation of fairness-aware policies. Through extensive experiments on cooperative use cases, we demonstrate how MAFE facilitates the design of equitable multi-agent algorithms and reveals critical trade-offs between fairness, performance, and coordination. MAFE provides a foundation for systematic progress in dynamic, multi-agent fairness research.

💡 Research Summary

The paper addresses a critical gap in algorithmic fairness research: most existing work focuses on static or single‑agent settings, while real‑world decision systems involve multiple interacting entities whose actions unfold over several stages and affect long‑term outcomes. To enable systematic study of fairness in such dynamic, multi‑agent environments, the authors introduce MAFE (Multi‑Agent Fair Environments), an open‑source benchmark suite compatible with standard MARL libraries.

MAFE extends the Dec‑POMDP formalism by adding two component functions for each agent: a reward component c(R) and a fairness component c(F). Rather than emitting a single scalar per timestep, these functions output vectors of raw counts (e.g., number of deaths, population size, region‑specific metrics). This “decomposable primitive” design gives researchers the flexibility to construct a wide range of derived fairness metrics—rates, disparities, temporal aggregates—after the fact, something impossible with pre‑aggregated reward signals.

Three concrete environments are instantiated: MAFE‑Health (insurance provider, hospital, central planner), MAFE‑Loan (admissions, fund disbursement, debt management), and MAFE‑Edu (admissions, scholarship allocation, academic support). All three draw on publicly available datasets (Lending Club, IPUMS, NCES, CDC) to sample realistic individual attributes and to fit regression models that map features to outcome probabilities. Controlled amplifications introduce structural disparities, ensuring that the simulated societies exhibit inequities similar to those observed in the real world.

The authors evaluate representative cooperative MARL algorithms (QMIX, VDN) and several fairness‑aware variants that weight the fairness component in the agents’ reward signals. Empirical results show that (1) incorporating fairness into the reward typically reduces short‑term performance metrics (approval rates, treatment success, graduation rates) but substantially narrows group‑level gaps over time; (2) clear role separation among agents (cost‑minimizing insurer, health‑maximizing hospital, equity‑focused planner) facilitates policy coordination and yields more interpretable trade‑offs; and (3) the decomposable primitives enable both step‑wise and cumulative fairness assessments, revealing dynamics that scalar‑only baselines miss.

Compared with prior work, MAFE uniquely combines heterogeneous agents, real‑world data, and flexible fairness diagnostics within a single framework. Existing long‑term fairness environments are either single‑agent, assume homogeneous agents, or rely on abstract toy models that cannot capture population‑level disparities. MAFE fills this void, offering a realistic testbed for fairness‑aware algorithm design across finance, healthcare, and education.

Limitations are acknowledged: the current experiments focus on cooperative settings, leaving competitive or negotiation‑driven scenarios unexplored; and the suite does not yet include modules for complex policy instruments such as legal regulations or social incentives. Future directions include extending MAFE to non‑cooperative Dec‑POMDPs, integrating human‑in‑the‑loop feedback, and building richer policy‑intervention interfaces.

In summary, MAFE provides the first comprehensive, open‑source platform for evaluating fairness in multi‑agent, multi‑stage decision processes, enabling researchers to quantify fairness‑performance trade‑offs, test equitable MARL algorithms, and move toward policy‑relevant, data‑driven fairness solutions.