10 cards. Moving information to the agent whose decision it changes.
Every card shows its question first. Reveal one at a time to test yourself, or
reveal all of them to use the deck as a revision sheet.
10 cards
Intuition
Why does partial observability create a need for communication between agents?
Intuition
One agent may know information that another agent cannot observe directly.
Sharing that information can help the other agent make a better decision.
Concept
When is information from one agent useful to another agent?
Concept
When receiving that information can change the other agent's decision in a way that improves the team's outcome.
Equation
What does this mean?
Ai=Xi×Mi
Equation
Agent i's action can be represented as both:
a task or environment action from Xi;
a communication action from Mi.
Communication can therefore be treated as part of the agent's decision.
Equation
What does this represent?
mti=fi(oti)
Equation
Agent i generates message mti from its current observation oti.
The function fi determines what information the agent communicates.
Equation
What does this tell us about a b-bit communication channel?
∣M∣=2b
Equation
A b-bit message can represent 2b distinct messages.
For example, 3 bits provide 8 possible messages.
Intuition
Why is sending more information not always better?
Intuition
Communication may have limited bandwidth, cost, noise, delay, or loss.
The goal is not maximum communication. It is useful communication.
Equation
What does this represent?
rt′=rt−λct
Equation
A reward that charges the team for the communication it used.
rt
the task reward
ct
communication usage
λ
how costly communication is
Concept
How can capacity, noise, loss, range, or cost change an agent's communication strategy?
Concept
Agents may need to communicate more selectively, compress information, avoid unnecessary messages, or learn strategies that remain useful when messages are unreliable.
Concept
What is a learned communication protocol?
Concept
A communication protocol is learned when agents develop useful relationships between information, messages, and behaviour through training, rather than having every message meaning specified in advance.
Scenario
Why can two successful communication protocols fail when agents trained with different conventions are paired?
Scenario
The agents may assign different meanings to the same messages.
A message only works if the sender and receiver share compatible conventions.