Parts 2 to 4: Coordinate, Communicate, Adapt

6 min read

The central project design combines three linked decisions. First, identify one coordination failure and select a training approach that addresses it. Second, specify useful message content, recipients, representation, and at least one communication constraint. Third, define a partner or environment change and choose a strategy for adaptation. Each choice must follow from the system formulation and state a behavioural consequence that the evaluation can later measure.

---

Part 2 · Coordinate

Identify one coordination problem

Name one, specifically. Examples of the right grain:

• multiple drones searching the same area while another region goes uncovered,
• two ground robots attempting the same rescue,
• agents blocking one another in a corridor,
• nobody maintaining the communication relay because every role is more urgent,
• a poor division of roles when several agents could do several jobs.

“Coordination is hard” is not a coordination problem. Neither is a list of five. Pick the one that would most damage your system and describe how it arises from the observations and actions you defined in Part 1.

Choose a training approach

Then answer:

How should your training approach encourage coordinated behaviour?

Choose and justify one:

Approach	Choose it when	Accept that
Independent learning	simplicity and scale matter; you need a baseline	non-stationarity, high variance across seeds
Centralized critic	learning is unstable because partner behaviour confounds the signal	one more component to train
Value decomposition	credit assignment is the bottleneck, and rewards are common	the mixer grows with agent count
Centralized execution	the team is small and genuinely co-located	an exponential joint action space, and no decentralized deployment

Key takeaway

The Challenge Lab is the argument for choosing rather than accumulating.

There, value decomposition improved reward by 0.06 while communication improved it by 1.43, on the same problem. Coordination machinery was the wrong tool because the bottleneck was missing information, not credit.

So the justification you need is not “value decomposition is good”. It is “the coordination problem in my system formulation is a credit assignment problem, and here is why”, or the equivalent for whichever you pick.

You do not have to use every technique in this resource. Saying which you rejected, and why, is worth more than using all of them.

---

Part 3 · Communicate

Four decisions, and Part 1 should already have told you what the first one is.

What should be communicated?

Look back at the informational awkwardness you identified in Part 1: the decision an agent cannot make well from its own observation. What would fix it?

Candidates in this domain: discovered survivors, hazards, an intended route, task status, remaining battery.

Apply the three tests from Message Content to each:

• Does the receiver already know it?
• Can the receiver act on it?
• Would it choose differently having heard it?

A message that fails any of the three is costing bandwidth for nothing.

Who receives it?

Broadcast to all, nearby agents only, or specified recipients such as relay agents. Say which, and note that range makes communication depend on the state.

Message representation

Discrete symbols, a small vector, or a structured status message. If discrete, say how many, because that is your capacity and it decides which situations must merge.

Choose at least one constraint

Constraint	What it forces you to think about
Bandwidth	which distinctions to keep and which to merge
Message loss	whether silence carries meaning in your protocol
Range	whether the protocol survives agents moving apart
Cost	whether each message is worth its price

Then answer:

What information is important enough to justify communication?

Common pitfall

The Challenge Lab’s communication result is the thing to bear in mind here.

The same four messages raised throughput from 7.62 to 7.70 and lowered the team reward from 7.56 to 7.43. The information was real and the price was higher, so the protocol was a net loss on the objective while looking like a win on the task measure.

Communication is not valuable in itself, and it is not valuable merely because it carries useful information. It is valuable when the change it produces exceeds what the channel costs. Say what your protocol would have to buy to justify its price, and if your Part 1 observations are broadly symmetric, consider whether your system needs a channel at all. That would be a defensible and unusually honest answer.

---

Part 4 · Adapt

Now deployment uncertainty. Your system must handle at least one unfamiliar condition involving other agents.

Choose the condition

• a new rescue drone joins mid-operation,
• one agent fails and stops responding,
• robots from another organisation arrive,
• an agent follows a policy you did not train,
• the number of active agents changes.

The last two are the harder and more interesting ones. The first three can sometimes be handled by ordinary robustness.

Training strategy

How will the system avoid depending on one fixed group of agents?

Options, with what each costs:

Strategy	Cost
Diverse partner policies	peak performance with any one partner
Population-based training	compute, and the work of building the population
Randomised roles during training	slower convergence
Behaviour variation across episodes	you must decide what to vary, and verify it varies

The warning from Training Partner Diversity applies: the number of partners is not behavioural diversity. Ten near-identical partners present one pattern. Say how you would check that yours actually differ.

Adaptation mechanism

Will your agent

• infer other agents’ behaviour from observation,
• maintain a partner representation,
• react to observed actions without identifying anyone,
• or rely on a shared communication convention?

Figure 1

$$\ubrace{observe}{\text{observed behaviour}}{what I have seen} \rightarrow \ubrace{policy}{\text{representation}}{who I think you are} \rightarrow \ubrace{action}{\text{my action}}{what I do about it}$$

$\text{observed behaviour}$

locally available evidence about another agent

$\text{representation}$

the inferred behavioural summary

$\text{my action}$

the policy response conditioned on that summary

An adaptation mechanism maps local behavioural evidence to a conditioned action.

Name the behavioural consequence

This is the part most often skipped, and it is where the marks are.

How would the system behave differently after observing the unfamiliar agent?

An actual difference in action, not a statement that the system “adapts”. Something of this shape:

After three steps in which the new drone repeatedly enters sector B, my agents’ partner representation shifts toward area-focused, and their policy reallocates from sector B to sector D, leaving the new drone to finish B alone.

Key takeaway

Two failure modes to avoid, both measured earlier in this resource.

Committing once and never revising. In the Adapt Lab, the adaptive agent identified its partner after four steps and never reconsidered, which made it the weakest of three arms when the partner changed mid-episode.

Assuming a stranger resembles someone known. The same agent scored 25 on a held-out partner near its training set and 2 on one outside it. A partner model fails confidently, not gracefully.

---

Integrated Design Check

You should now have exactly:

• one coordination problem, and one training approach that addresses it
• one communication design: content, recipients, representation, constraint
• one unfamiliar condition, a training strategy, an adaptation mechanism, and a concrete behavioural consequence

If you have several of each, cut. A design that commits is assessable; a survey is not.

Next: Part 5.