MAMO multi-agent system: multi-objective constrained optimization

MAMO, a Multi-Agent system for Multi-Objective constrained optimization, is a proposal to separate task execution from objective design by learning reward weights rather than hand-tuning them. The paper, by Federica Filippini, was submitted to arXiv on 18 June 2026 as arXiv:2606.20236 and presented at the 17th Workshop on Optimization and Learning in Multiagent Systems (OptLearnMAS), co-located with AAMAS 2026.

What is MAMO and what problem does it target?

MAMO addresses cost-minimization problems under performance constraints in dynamic environments by using multi-agent reinforcement learning to choose reward weights. Traditional approaches embed costs and constraint violations into a single scalar reward using weighted penalty terms in a Lagrangian-inspired formulation, and the resulting policy behavior depends critically on those manually selected weights. MAMO treats the selection of those weights as a learning problem, decoupling objective design from task execution so agents can adapt when the environment changes.

The paper frames the core difficulty as the manual choice of penalty weights, which makes it hard to balance the primary objective and constraint avoidance, especially under non-stationary conditions. MAMO proposes multi-agent RL as the mechanism to learn that balance at runtime rather than relying on fixed, preselected weights.

How does MAMO differ from single-agent, penalty-weight approaches?

MAMO replaces a single-agent policy that optimizes a pre-weighted scalar reward with a multi-agent setup where one part of the system executes tasks and another part learns the reward weights. The abstract contrasts the common practice of embedding costs and violations into one reward with weighted penalties, against MAMO's approach which formulates reward-weight selection as an explicit learning objective.

The key distinction is procedural: the common Lagrangian-inspired method requires manual tuning of weights and thus hard-codes the trade-off between objectives, whereas MAMO aims to let learning determine those trade-offs. The paper positions this as a first step toward more autonomous and robust RL-based solutions for constrained optimization problems in dynamic environments.

Why it matters

MAMO targets a recurring operational problem in applied RL: selecting penalty weights that control safety or constraint adherence while preserving task performance. Manual tuning imposes deployment friction and can break down under shifting task priorities or changing environmental dynamics. By moving weight selection into the learning loop, the approach could reduce laborious hyperparameter tuning and make RL policies more resilient to non-stationarity, which matters for networking, computing systems, and other domains the paper cites as naturally formulable as cost-minimization under constraints.

The submission and presentation context also matters: the work was shared at OptLearnMAS, co-located with AAMAS 2026, signalling the paper was positioned for a community focused on optimization and multiagent learning rather than as a purely theoretical RL contribution.

What to watch

Check the paper's arXiv entry (arXiv:2606.20236) for follow-up artifacts: the page lists Code, Data and Media sections where practical evaluations or code could appear. The next concrete signals that will validate MAMO are public code or experimental results demonstrating learned weight selection improving constraint adherence in changing environments, and any peer-reviewed proceedings from OptLearnMAS/AAMAS that include the workshop presentation.

References and source facts pulled from the paper's arXiv metadata: title, author Federica Filippini, submission date 18 June 2026, arXiv identifier 2606.20236, and the comment that the work was presented at the 17th Workshop on Optimization and Learning in Multiagent Systems (OptLearnMAS), co-located with AAMAS 2026. The abstract describes the methodological shift from manual weighted penalties in Lagrangian-inspired rewards to treating reward-weight selection as a learning problem via multi-agent RL.