Notebooks
A browser lab and three lightweight notebooks provide the computational
activities for coordination, communication, adaptation, and the wireless
Challenge Lab. None requires a MARL framework, a GPU, or a large training job.
The environments, baselines and plots come from the
cooperative-marl-labs package, so a learner can
change conditions, compare policies and interpret results on a free Colab CPU
rather than spend the lab implementing an environment.
Coordinate: a virtual lab, not a notebook
Section titled “Coordinate: a virtual lab, not a notebook”The Coordinate practical runs in the browser, not in Colab. It is the lowest-friction practical in the resource: nothing to open, nothing to install, and you can drive the agents yourself before any learning happens.
- Take joint actions by hand on a two-agent switch grid.
- Watch the joint action space grow as agents are added.
- Train independent learners and value decomposition in the page.
- Compare their curves, metrics and replayed behaviour.
Open the Coordinate virtual lab
Communication Lab
Section titled “Communication Lab”
Two agents in the smallest task where communication can matter: one sees which dish was ordered, the other is the only one who can act.
Six experiments, and what each one establishes:
| Experiment | Result |
|---|---|
| No usable channel | 0.25, which is chance over four dishes |
| Four symbols | 1.00, learned with no dictionary |
| Capacity sweep | success is exactly : a narrow channel merges situations |
| Message loss | degrades as , the signature of a loss-oblivious protocol |
| Read the protocol | the learned mapping is a permutation, not the identity |
| Pair with a stranger | cross-play falls from 1.00 to 0.300, some pairings below chance |
Accompanies Communication Lab.
Adapt Lab
Section titled “Adapt Lab”One kitchen with collisions, one learning agent, six possible partners. Four available for training, two held out: one behaviourally near the training set and one outside it.
Compares three arms, and produces three findings that were not engineered:
- a specialist trained with one partner scores zero with a perfectly competent one it never met,
- that specialist beats the generalist on a stranger resembling its own partner,
- and it has the smallest generalization gap while being the worst agent, because its familiar-partner average is dragged down by a failure.
Accompanies Adapt Lab.
Challenge Lab: Cooperative Wireless Resource Allocation
Section titled “Challenge Lab: Cooperative Wireless Resource Allocation”Four access points, three channels, and no kitchen. The transfer task: the problem is stated in its own vocabulary and nobody maps the concepts for you.
Ten sections, ending in a comparison table you assemble and one design decision you defend. Three results worth knowing about in advance:
A deterministic local rule scores below random. Greedy channel selection reaches 3.37 against random’s 6.00, because four access points running the identical rule on similar observations move to the same channel together.
Communication raises throughput and lowers the team reward. From 7.62 to 7.70 on the task measure, and from 7.56 to 7.43 on the objective, because four messages at 0.05 each cost more than the information was worth. Useful is not the same as worth it.
The talking system does not degrade under a traffic shift, it collapses. It falls to 3.62, below random, by putting all four access points on one channel every step. Individual state coverage cannot explain it: 0.0% of those decisions use an unseen state. What was never trained is the joint configuration.
Accompanies The Challenge.
Running them
Section titled “Running them”Each notebook opens with one install cell:
!pip install -q cooperative-marl-labsEverything the labs need is in that package, so a
notebook holds the experiment rather than the environment. The Communicate lab
adds the PyTorch extra, cooperative-marl-labs[learning], because its protocol
experiments train a small network.