Skip to content
MARL in Cooperative Environments
Edit this page

2.1Communicate

2 min read

Two agents are preparing meals together.

  • Agent A can see the incoming orders.
  • Agent B can see the stove.
  • Agent A sees that soup is needed next.
  • Agent B sees that the current dish is almost finished.
  • Neither agent has the full picture.

Agent A now knows something that could change Agent B’s next action.

sees the order

sees the stove

Different information. Shared task. Communication connects the missing pieces.

  • When agents have different observations, useful information is distributed across the team.
  • Communication gives agents a way to share what could improve another agent’s decision.
  • The information is not missing from the system. It is in the wrong place.

That creates new questions:

  • What should be sent?
  • When should it be sent?
  • Who needs it?
  • How much information can the channel carry?
  • What happens when messages are noisy or lost?
  • What happens when agents learn different communication conventions?
  • Messages: Learn how communication can be represented as part of an agent’s action.
  • Message content: Explore what agents may share, including observations, intentions and plans.
  • Communication constraints: Study limited capacity, range, noise and message loss.
  • Communication policies: Understand how agents decide what to communicate, when to communicate, and when to remain silent.
  • Learned communication protocols: See how agents can develop their own conventions through learning.
  • Interpretable communication: Explore why an effective learned protocol may still be difficult for humans or unfamiliar agents to understand.
  • LangGround: Connect these ideas to recent work on grounding learned communication in human-interpretable language.
  1. Private information
  2. Message
  3. Shared information
  4. Better joint decisions

By the end of this chapter, you should be able to reason about when communication is useful, what information should be transmitted, and why a successful learned protocol may still fail with unfamiliar agents.