Research map / Multi-agent systems and games / Games and equilibria
Beyond Mandatory Federations: Balancing Egoism, Utilitarianism and Egalitarianism in Mixed-Motive Games
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Flexible-Participation Federation (FPF) and Local Multi-Federation (LMF) so agents can voluntarily join federations, balancing egoism, utilitarianism and egalitarianism in mixed-motive games with convergence guarantees.
Problem
Mixed-motive multi-agent learning methods require agents to share private local information or mandatorily join federations, which is impractical and risks free-riding and exploitation.
Approach
Replace mandatory global federation with voluntary participation (FPF), where participating agents regularize their egoistic policy toward a federated utilitarian/egalitarian policy, and extend to LMF where agents form multiple localized federations exchanging parameters or policy gradients.
Key results
- FPF and LMF achieve egalitarianism (low std/mean reward ratio) comparable to mandatory-participation D&F in Matthew Effect and Manufacturing Plant scenarios.
- Agents opting out of FPF federation experience lower average returns than participating agents across Job Scheduling, Matthew Effect and Manufacturing Plant.
- Theory: the global FPF model achieves an O(1/T) convergence rate; LMF agents reach consensus with gap O(1/T^beta + epsilon_bar^T).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FPF or LMF instead of the compared baselines | FPF and LMF · Cleanup and Harvest | collective returns | Helped |
| use FPF or LMF instead of the compared baselines | FPF and LMF · Classic Discrete Scenarios | collective returns | Mixed |
| use FPF or LMF instead of the compared baselines | FPF and LMF · Cleanup and Harvest | standard variance / mean of rewards | Helped |
| use FPF or LMF instead of the compared baselines | FPF and LMF · Classic Discrete Scenarios | standard variance / mean of rewards | Mixed |
| use LMF instead of D&F | LMF · MPE Predator-prey | collective returns | Helped |
| use LMF or D&F instead of other compared methods | LMF and D&F · MPE | collective returns | Helped |
| opt out of federation participation | FPF agents · Job Scheduling, Matthew Effect, and Manufacturing Plant | individual returns | Hurt |
| use full-participation LMF instead of partial-participation FPF | LMF and FPF · Classic Discrete Scenarios | collective returns | Helped |
| form multiple joint federations dynamically instead of fixed disjoint federations (ablation) | LMF · Matthew Effect | collective returns | Hurt |
Surprising findings
- In Job Scheduling, IPPO-AVG and SOTO outperform FPF and LMF on egalitarianism, which the authors attribute to the single shared resource limiting penalization of greedy agents without reward shaping.
- LMF-Fix (two disjoint federations) achieves better utilitarian performance than LMF-Dynamic (multiple joint federations), as overlapping federations complicate consensus.
Limitations
- Free entry and exit of agents during training is not supported; future work mentions it.
- No robust metric yet for exploitation risk of participating agents in federations.
- FPF under partial participation underperforms full-participation LMF and mandatory D&F on utilitarian returns in some settings.
Experimental setup
Datasets: Job Scheduling Matthew Effect Manufacturing Plant Cleanup Harvest Multi-Agent Particle Environments (Physical deception, Covert communication, Predator-prey, Keep-away)
Benchmarks: Job Scheduling Matthew Effect Manufacturing Plant Cleanup Harvest Physical deception Covert communication Predator-prey Keep-away
Metrics: collective returns standard deviation / mean ratio of agent rewards average returns of participating vs non-participating agents
Baselines: IDQN IPPO IPPO-AVG QMIX MAPPO MADDPG FEN SOTO-GGF SOTO-ALF D&F Inequality Aversion Social Influence
Builds on: D&F framework (Dong et al. 2024b) Inequality Aversion SOTO FEN
Closest papers by meaning
- Learning to Balance Altruism and Self-interest Based on Empathy in Mixed-Motive Games (NeurIPS 2024 · similarity 0.89)
- Aligning Individual and Collective Objectives in Multi-Agent Cooperation (NeurIPS 2024 · similarity 0.88)
- Fair Cooperation in Mixed-Motive Games via Conflict-Aware Gradient Adjustment (NeurIPS 2025 · similarity 0.87)
- Training Generalizable Collaborative Agents via Strategic Risk Aversion (NeurIPS 2026 · similarity 0.85)
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.85)
- A Unified Model of Direct and Indirect Reciprocity in Multichannel Games (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.