Skip to content
MARL in Cooperative Environments
Edit this page

2.7Learning Communication Protocols

6 min read

A communication protocol assigns behavioural meaning to messages: a sender chooses a signal and a receiver changes its action in response.

In this section you will

  • Read a hand-designed message semantics
  • Describe a protocol learned from the shared task reward
  • Explain meaning through the effect a symbol has on the receiver

The straightforward option: a human fixes the vocabulary.

SymbolMeaning, assigned by us
0no request
1get an ingredient
2begin cooking
3serve the dish

Now the agents have one job left: learn when to use each symbol, and how to respond. The semantics are given, so the receiver’s side is partly solved before training starts. Send 1 and the receiver already knows it concerns ingredients.

This is a perfectly respectable engineering choice. It is inspectable, it is debuggable, and where you genuinely know what needs saying it is hard to beat.

Its limit is that you have to know. A hand-designed vocabulary can only express distinctions its designer thought of, and the useful distinction in a task is often not obvious in advance. This is the same objection as hand-coding a policy: it works exactly as far as your understanding of the problem does.

Now remove the meanings and keep the symbols.

Figure 1
M={0, 1, 2, 3}\tone{comm}{\mathcal{M}} = \{0,\ 1,\ 2,\ 3\}
M\mathcal{M}
the discrete message space
0,1,2,30,1,2,3
symbols whose meanings are learned through their effects
A discrete message space supplies symbols but no predefined semantics.

At the start of training, 2 means nothing at all. It is an integer the sender may emit and the receiver may condition on, and that is the whole of its definition.

What can happen during learning is this. The sender occasionally emits 2. On some of those steps the receiver, also learning, happens to respond in a way that works out. The team return goes up, both policies are reinforced, and the pairing of this situation with this symbol with that response becomes slightly more likely. Repeat, and a regularity forms.

The textbook is explicit about why this is hard: because agents do not initially know the meaning of communication actions, they have to learn how to interpret the messages of others. Sender and receiver are learning against each other, so neither side can be correct first.

Suppose the regularity forms. Here is what it looks like, and what it does not mean.

SituationSymbol sentSituationSymbol sent
ingredient needed3ingredient needed2
ingredient needed0ingredient needed2
ingredient needed2ingredient needed2
early in trainingno stable mappinglatera convention

The left column is noise: the same situation produces different symbols, because nothing has settled. The right column is a protocol: a stable mapping from situations to symbols, shared by sender and receiver.

Now the sentence worth keeping from this section.

Two consequences, and both matter later.

The labels are arbitrary. Nothing distinguishes the protocol above from one where 1 plays the role of 2. Any permutation of the symbols works equally well, provided both agents permute together.

The meaning lives in the pair, not in either agent. There is no fact about agent 1 that makes 2 mean anything. The meaning is a property of the relationship, and it exists only while both halves agree.

That second point is where this chapter starts turning into the next one. A protocol is a convention held jointly by the agents that learned it, and a convention held jointly is exactly the kind of thing that fails when one partner is replaced.

Knowledge check

A trained team uses symbol 2 whenever an ingredient is needed. You retrain the receiver alone, from scratch, keeping the sender frozen. What happens to the meaning of 2?

Select one answer.

  • Hand-designed messages fix the semantics, so agents only learn when to use each symbol. Inspectable and debuggable, and limited to distinctions the designer anticipated.
  • Learned messages start meaningless. The symbol space is given; the meanings are not.
  • Learning them is a bootstrapping problem: a message is useful only if the receiver responds usefully, and the receiver learns that only from useful messages.
  • A protocol is a stable mapping from situations to symbols, shared by sender and receiver.
  • A symbol means whatever its listeners do about it. The labels are arbitrary, and the meaning is a property of the pair rather than of either agent.