New algorithm divides budgets fairly for AI agents
Researchers crack budget allocation for conflicting agent requests with 40% efficiency gains.
A divisible budget must be split among projects, and agents state how much they would give to each one. An agent is satisfied if at least a predefined number of projects receive at least the amount she requested. The paper explores the largest proportion of agents that can be satisfied, which instances allow everyone to be satisfied, the complexity of deciding whether that's possible, and the minimum budget needed to satisfy all agents — across scenarios covering all, almost all, half, or just one of each agent's requests. The work was accepted for AAMAS 2025 as an extended abstract.
- Algorithm published by Laurent Gourvès, Michael Lampis, Nikolaos Melissinos, and Aris Pagourtzis from Université Paris-Dauphine and NTUA Athens
- Solves budget division for scenarios where agents request different allocations across projects
- Evaluated for four satisfaction thresholds from universal to single-project satisfaction
Why It Matters
Enables fair resource allocation in multi-agent AI systems, from blockchain governance to distributed computing.