Skip to content
MARL in Cooperative Environments
Edit this page

2.10Communication Failures

5 min read

Communication can fail even when a channel exists and every agent is trained successfully.

In this section you will

  • Distinguish missing information from irrelevant information
  • Distinguish a lost message from an ambiguous one
  • Recognize a protocol mismatch between agents
  • Match each symptom to the intervention that can actually repair it

The sender holds something useful and does not transmit it.

Agent 1 can see the order ticket and never mentions the dish. The information exists in the system, the channel exists, and the two are not connected.

How to spot it: compare each agent’s private information against what it sends. Anything the receiver could act on and never hears is a candidate.

Where it comes from: usually a communication policy that was never pressured to find the useful message, either because λ\lambda was high enough to discourage sending anything, or because the bootstrapping problem from Learning Communication Protocols never resolved.

Messages are sent and change no decision.

Agent 1 reports the pot temperature, which agent 2 can already see. Or reports the chopping board, which agent 2 cannot use. Bandwidth is spent, nothing changes.

How to spot it: ablate. Replace a message with ∅\varnothing and see whether performance moves. If it does not, that message was not doing anything.

Why it matters even when it seems harmless: it costs λct\lambda c_t every step, it makes the protocol unreadable by burying the useful part, and under a narrower channel it displaces something that mattered.

The protocol assumes reliable delivery and messages disappear.

Communication Constraints covered the mechanism. The sharp version is that a lost message becomes ∅\varnothing, which is a legitimate message meaning “say nothing”, so a protocol in which silence carries meaning misreads loss as content.

How to spot it: sweep plossp_{\text{loss}} and look at the shape of the decline. A protocol robust to loss degrades gently. One that depends on silence falls off a cliff.

The message arrives and does not narrow the receiver’s options enough.

Agent 1 sends 1, meaning “help needed”. Agent 2 does not know whether to fetch an ingredient, take over the stove, or start plating. It has been told that something is wrong and not what.

How to spot it: look for messages after which the receiver’s action distribution stays wide. If receiving a message leaves the receiver as uncertain as before, it did not resolve anything.

Where it comes from: most often a capacity limit. With two messages available, several situations must share one symbol, and the receiver has to disambiguate from context. That is a design choice with a known cost, not necessarily a bug.

Two individually successful agents interpret the same message differently.

Agent A learned 2 means bring an ingredient. Agent D learned 2 means serve the dish. Both scored highly with their own partners. Paired, they fail, and no amount of retraining either one alone identifies the fault.

How to spot it: cross-play. Pair policies that were not trained together and compare against their within-team scores. A large drop is the signature. the Adapt chapter makes this a measurement discipline.

Why it has no single-agent analogue: there is no such thing as a convention held by one agent. The failure lives in the relationship, so it is invisible to any evaluation that tests agents only with the partners they trained with.

FailureWhat is wrongFirst thing to try
Missing informationuseful facts never sentcheck private information against messages sent
Irrelevant informationmessages change no decisionablate messages and watch performance
Message lossprotocol assumes deliverysweep plossp_{\text{loss}}
Ambiguitymessage does not narrow optionscheck the receiver’s action spread after a message
Protocol mismatchnothing, in either agentcross-play against unfamiliar partners

Knowledge check

A team scores 94% together. Each agent, paired with an equally strong agent from a different training run, scores 31%. Which failure is this?

Select one answer.

  • Missing information: useful private facts are never sent.
  • Irrelevant information: messages arrive and change no decision. Ablate to detect.
  • Message loss: the protocol assumed delivery. Worst where silence carries meaning.
  • Ambiguity: the message does not narrow the receiver’s options, usually a consequence of limited capacity.
  • Protocol mismatch: two working protocols disagree. Nothing is wrong with either agent, the failure lives in the relationship, and only cross-play reveals it.