2.5Communication Constraints

5 min read

By the end of this section you can answer

What happens to a protocol when the channel is limited?

Communication channels constrain which information agents can exchange and how reliably it arrives.

In this section you will

• Reason about a finite message capacity
• Reason about a limited communication range
• Reason about messages that are dropped
• Reason about messages that arrive corrupted

Message Capacity

A discrete message space has a size, and that size is a hard limit on how many distinct things can ever be said.

Figure 1

$$\bigl|\tone{comm}{\mathcal{M}}\bigr| = K \qquad\text{or, for a } b\text{-bit message,}\qquad \bigl|\tone{comm}{\mathcal{M}}\bigr| = 2^{b}$$

$K$

how many distinct messages exist

$b$

bits available per message; each bit doubles the space

Capacity counts distinctions, not detail. K messages can express exactly K situations.

Drag the slider to see what a bit buys.

Bits per message

Capacity is best thought about backwards, starting from the distinctions the team actually needs. With one bit, agent 1 has exactly two things it can say:

Message	Meaning
0	no help needed
1	help needed

That is a real protocol and it is often enough. But notice what it cannot do: it cannot say what kind of help. Two bits can.

Message	Meaning
00	no help needed
01	bring an ingredient
10	start cooking
11	ready to serve

Key takeaway

A narrow channel does not force a team to say less about each situation. It forces the team to merge situations.

With one bit, “bring an ingredient” and “start cooking” have to become the same message, and the receiver has to work out which from context. Designing a protocol under a capacity limit is deciding which distinctions are worth keeping.

Communication Range

A message may reach only agents that are close enough.

In the Dec-POMDP this is part of the observation model rather than a separate mechanism: whether agent $j$ observes $m^\ag_t$ depends on the state, through the distance between them.

Figure 2

$$\text{agent } j \text{ receives } \tone{comm}{m^\ag_t} \iff \mathrm{dist}\bigl(\ag,\ j\bigr) \leq \ubrace{observe}{d}{the range}$$

$m^i_t$

the message sent by agent i at step t

$\mathrm{dist}(i,j)$

the current distance between sender and receiver

$d$

the maximum communication range

Message receipt depends on the state-dependent distance between agents.

Two consequences worth noticing. Agents cannot rely on being heard, so a protocol built on “everyone knows what I said” fails as soon as someone walks away. And range makes communication state-dependent: the same message sent from a different position reaches a different audience.

Message Loss

A sent message may simply not arrive.

The same picture with an unreliable channel. Agent 1 sends; agent 2 may receive nothing at all.

The standard way to model this is with a probability: with probability $p_{\text{loss}}$, the message the receiver observes is replaced by the empty message.

Figure 3

$$\text{received} = \begin{cases} \tone{comm}{m^\ag_t} & \text{with probability } 1 - p_{\text{loss}} \\[2pt] \tone{conflict}{\varnothing} & \text{with probability } p_{\text{loss}} \end{cases}$$

$\varnothing$

the empty message: the receiver gets nothing, exactly as if none had been sent

Loss turns a message into a silence, and silence is a message the sender did not choose.

Loss is worse than it first appears, because $\varnothing$ is a legitimate message that means “say nothing”. So a receiver cannot distinguish

• the sender chose to stay silent, from
• the sender spoke and the channel ate it.

A protocol in which silence carries meaning is therefore fragile under loss. If no message means “carry on as planned”, then a dropped “stop” reads as approval.

Communication Noise

Weaker than loss and sometimes more damaging: the message arrives, changed.

For a continuous message the usual model adds a random perturbation.

Figure 4

$$\text{received} = \tone{comm}{m^\ag_t} + \tone{conflict}{\epsilon}, \qquad \epsilon \sim \mathcal{N}(0, \sigma^{2})$$

$\epsilon$

noise added by the channel

$\sigma^{2}$

how much of it

The message arrives, slightly wrong. Nothing tells the receiver by how much.

Noise punishes a particular kind of protocol: one where nearby messages mean very different things. If a learned vector uses tiny differences to distinguish “bring a tomato” from “bring an onion”, noise destroys it. A protocol whose meanings are far apart in message space survives.

Key takeaway

A useful communication strategy should remain useful under the constraints of its channel.

A protocol is not just a mapping from situations to symbols. It is a mapping that has to work through a specific channel, with that channel’s capacity, reach, reliability and noise.

A protocol tuned on a perfect channel and deployed on a real one is one of the most common ways cooperative systems fail.

Example

A kitchen protocol can face all four constraints. A small symbol set limits capacity, only nearby agents may hear a signal, equipment noise can hide or corrupt it, and sending a message can consume time. Silence is therefore an action worth considering rather than proof that communication is absent.

Knowledge check

A team's protocol uses silence to mean 'continue as planned'. What goes wrong when 10% of messages are lost?

Select one answer.

AA lost message becomes silence, so a dropped instruction is read as approval to continue.BNothing much. Losing 10% of messages costs at most 10% of performance.CThe receiver will notice the gap and ask again.DThe messages arrive corrupted rather than missing.

Check answer

Communication Constraints Summary

• Capacity $|\mathcal{M}| = K$, or $2^b$ for $b$ bits, counts distinctions. A narrow channel forces the team to merge situations, not to describe them more briefly.
• Range makes communication state-dependent: who hears you depends on where everyone is.
• Loss replaces a message with $\varnothing$, which is indistinguishable from a chosen silence. Protocols in which silence means something are fragile.
• Noise perturbs the message, punishing protocols whose meanings sit close together in message space.
• The principle: a protocol must remain useful under its own channel’s constraints, not just on a perfect one.