← News

Agents & MCPSources: arXiv · 16 Nov 2026

Axiomatic framework ties reinforcement learning in open multi-agent systems to Shapley axioms

Researchers propose deriving value function learning from the Shapley axioms in n-agent ad hoc teamwork, where only some agents are controllable, and report consistent gains on NAHT benchmarks.

In plain words

In many real systems, such as power grids, a learning method can direct only some of the participating agents, not all of them. The authors build a rule-based foundation for this setting by treating three classic fairness axioms from cooperative game theory as hard requirements for how agents' values are learned. Following those rules, they say, leads to algorithms that learn more efficiently and handle unfamiliar teammates better.

What was announced

The paper, an arXiv abstract, addresses open multi-agent systems in which only a subset of agents are controllable, a setting the authors say has lacked a principled foundation for reinforcement learning algorithm design. They study it through n-agent ad hoc teamwork (NAHT) and propose an axiomatic framework that derives value function learning from the Shapley axioms of Efficiency, Symmetry and Linearity, treating them as structural constraints rather than heuristic choices. The framework rests on a new NAHT decomposition that the authors say characterises the difference between open and fully controllable systems.

According to the abstract, enforcing the axioms on individual value functions recovers the Shapley value for Dec-POMDPs, which the authors describe as a bridge between reinforcement learning and classical cooperative game theory. They introduce two algorithmic templates, Shapley Machines and Banzhaf Machines, that can be instantiated with standard reinforcement learning methods. Applied to base algorithms such as IPPO and POAM, the authors report consistent improvements in learning efficiency and in generalisation to unseen agent types and conventions across NAHT benchmark tasks, with experiment code published on GitHub.

What it does not show

  • The evidence is the paper's own abstract; the reported improvements and theoretical claims are not independently verified, and no benchmark numbers are given.

For security and IT teams

Scope
The source is an arXiv abstract (v3, replace-cross announcement) covering the framework, the two algorithm templates and reported benchmark behaviour, not the full paper's proofs or experiments.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.