The Common Information Approach to Decentralized Coordination

This direction is split into other directions: common information coordination. Its last version stays readable and citable here.

Version 1 of 1105 papers161 conceptsfrozen 2026-10-04

In a centralized system, a single "brain" sees everything and decides everything. But in the real world—whether it's a swarm of drones, a power grid, or a team of players in a card game—information is scattered. Each agent sees a different slice of the world. This asymmetry creates a computational nightmare: to act optimally, Agent A must reason about what Agent B knows, what Agent B thinks Agent A knows, and so on, into an infinite regress of nested beliefs.

The Common Information Approach provides a mathematical escape hatch from this complexity. By reframing decentralized problems through the lens of shared knowledge, researchers have found a way to collapse these infinite hierarchies into a single, manageable "public" state.

The Virtual Coordinator and the Power of Prescriptions

The core insight, pioneered by researchers like Aditya Mahajan, Ashutosh Nayyar, and Demosthenis Teneketzis, is to imagine a fictitious coordinator. This coordinator doesn't see any agent's private data; it only sees the information that is common knowledge to everyone (such as the history of public actions or shared sensor data).

Instead of telling an agent exactly what to do, the coordinator issues a prescription: a mathematical function that maps an agent's private observations to a specific action.

As explained in Decentralized Stochastic Control with Partial History Sharing: A Common Information Approach, this shift from "choosing actions" to "choosing functions" is transformative. It allows the system to maintain local autonomy—agents use their private data to make the final call—while ensuring global coordination. The coordinator’s job is to find the best set of functions to hand out, effectively turning a messy decentralized problem into a centralized one.

Collapsing Beliefs into a Public State

Why does this work? In a standard Decentralized POMDP, the "state" is a moving target because agents are constantly updating their beliefs about each other. However, Aditya Mahajan and Ashutosh Nayyar demonstrated that the common information state—the probability distribution over the global state and everyone's private data, conditioned on shared history—acts as a sufficient statistic.

This means that all the relevant "history" of the system can be compressed into a single belief state that everyone agrees on. As Serdar Yuksel notes in A Universal Dynamic Program and Refined Existence Results for Decentralized Stochastic Control, this transformation restores a "Markovian flow" to the system. We can now use standard tools like Dynamic Programming to solve problems that were previously considered intractable.

From Control Theory to Deep Learning

While these ideas originated in control theory, they have recently revolutionized Multi-Agent Reinforcement Learning (MARL). A prime example is the Bayesian Action Decoder (BAD), developed by Jakob Nicolaus Foerster and colleagues in Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning.

BAD applies the common information logic to the cooperative game Hanabi. In Hanabi, players see everyone's cards except their own. BAD agents maintain a public belief about their own private cards based on the actions others have taken. By acting in the space of "partial policies" (prescriptions), the agents can intentionally take actions that reveal their private information to the team, discovering complex communication conventions without any pre-defined language.

Scaling to Large Populations

The common information approach also provides a path toward controlling massive groups of agents. In Team optimal control of coupled subsystems with mean-field sharing, Jalal Arabneydi and Aditya Mahajan show that if agents share a "mean-field" (the average state of the population), the complexity of finding an optimal strategy becomes independent of the number of agents.

The global average becomes the "common information," allowing a virtual coordinator to prescribe a single rule that every agent follows locally. This makes it possible to coordinate thousands of electric vehicles or robots with the same mathematical rigor used for a two-agent system.

Summary of the Intuition

The Common Information Approach works because it identifies a shared perspective in a world of private views. By focusing on what is common knowledge:

  1. We replace the infinite regress of "I think that you think" with a single public belief.
  2. We transform decentralized agents into local executors of centralized prescriptions.
  3. We reduce NEXP-complete decentralized problems into solvable centralized MDPs.

This framework, as explored across decades of work by researchers like Aditya Mahajan and Deepanshu Vasal, provides the theoretical bridge needed to move from "independent agents" to "truly cooperative teams."

Papers in this version

Showing 60 of 105 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_common_information_approach_to_decen_e08931ec,
  title = {The Common Information Approach to Decentralized Coordination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-common-information-approach-to-decentralized-coordination/rdr_e08931ec30c04ae6a864098c3ad24738},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_common_information_approach_to_decen_e08931ec_v1,
  title = {The Common Information Approach to Decentralized Coordination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-common-information-approach-to-decentralized-coordination/rdr_e08931ec30c04ae6a864098c3ad24738/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}