Deliverables

4 min read

The final submission consists of three connected artifacts: a system diagram, a concise design brief, and an evaluation plan. This section specifies the content of each artifact, identifies work that is outside the required scope, and describes an optional implementation extension. Together, the artifacts should show what each agent knows and does, why the selected mechanisms address the stated cooperative problems, and how the resulting claims would be tested.

1 · System diagram

One diagram showing

• your agents, by role,
• the environment,
• what each role observes,
• what each role does,
• the communication links, and who hears whom.

Hand-drawn and photographed is fine. Boxes and arrows are fine. What matters is that a reader can see which information reaches which agent, because that is the claim the rest of the brief rests on.

One thing to get right

Draw the observation arrows from the environment to each agent individually, not as one arrow to a box labelled “agents”.

A single shared arrow is how designs quietly become centralized. If your diagram cannot show that two agents receive different things, it is not showing a decentralized system.

2 · Design brief

Maximum two pages. Five short sections, matching the five parts.

Section	Content
System	agents, observations, actions, reward, objective
Coordinate	the one coordination problem, the training approach, why that approach for that problem
Communicate	what is sent, to whom, in what representation, under which constraint
Adapt	the unfamiliar condition, training strategy, adaptation mechanism, and the behavioural consequence
Evaluate	metrics and test conditions, in prose or pointing at artifact 3

Two pages is a real constraint and it is there on purpose. Fitting a complete design into it requires having decided things.

Common pitfall

What loses marks in this brief, in order of frequency:

Naming mechanisms without connecting them to problems. “We use CTDE with a centralized critic and value decomposition” says nothing unless the previous paragraph named the difficulty each one addresses.

An observation that is really the state. If an agent knows where the survivors are, there is no search problem.

Adaptation with no behavioural consequence. “The system adapts to new agents” is a claim, not a design. Say what an agent does differently.

A reward with eight terms. Each additional term is another thing that can be maximised without doing the task.

3 · Evaluation plan

One compact table. The experiments you would run, and what each one is for.

Test condition	Metric	What it would establish
Familiar agents	survivors reached, duplicated coverage	baseline, and whether coordination is real
Held-out agent	survivors reached, time to first assist	partner generalization
30% message loss	survivors reached, messages sent	communication robustness
One agent fails at $t = 50$	coverage, response time	adaptation
Timestep-only policy	survivors reached	whether the task requires coordination at all

That last row is the falsification test, and including something like it is the single clearest signal that you understood Part 5.

Excluded Work

No implementation. You are not asked to train anything.

No report. The two-page brief is the report.

No literature review. Cite something if it helped you decide, and do not survey.

No use of every technique. A design using independent learning, defended, beats one listing four methods without reasons.

Optional Implementation Extension

For anyone who wants to go further. Explicitly optional, and the project is complete without it.

Implement a simplified version of one component:

• simulate communication loss and measure how your protocol degrades,
• build a small partner classifier and check what it does with a stranger,
• create a minimal grid world with two of your roles,
• compare two candidate reward functions and show they produce different behaviour.

Key takeaway

If you do the extension, the most valuable thing you can produce is a negative result about your own design.

Finding that your protocol collapses at 20% message loss, or that your partner classifier labels every stranger as the same familiar type, is worth more than a plot showing your system working. Both labs in this resource were built around findings of exactly that kind.

The notebooks in this resource are reasonable starting points. The Adapt chapter lab is about a hundred lines of pure Python and demonstrates partner generalization end to end, which is a realistic scale for an optional extension.

Submission Package

Save or submit the three artifacts together. A single PDF with the diagram, the brief and the table is ideal.

The Rubric lists what each of the four criteria is looking for, and it is worth reading before you write rather than after.