3.13Adapt Worksheet

3 min read

This five-part digital activity consolidates the chapter’s approach to partner generalization. You will match adaptation concepts to their applications, calculate a familiar-to-unseen generalization gap, and analyze a cross-play table. An interactive graph compares familiar and held-out partner performance for specialist, generalist, and adaptive agents. The final prompt asks you to weigh specialization against robust deployment performance. Every earlier item is automatically checkable.

What you will be able to do

• Match partner-generalization concepts to their use in cooperative MARL. Understand
• Calculate simple measures of familiar and unseen-partner performance. Apply
• Interpret results across different partners. Analyse
• Evaluate the trade-off between specialization and partner generalization. Evaluate

1 · Match Concepts to Applications

Drag each block into a slot, or tap one and then tap a slot.

• Partner diversity
• Agent modelling
• Ad hoc teamwork
• Zero-shot coordination
• Cross-play

1. Train with several different partner policies.

   drop here

2. Infer how another agent is likely to behave.

   drop here

3. Cooperate with previously unknown agents.

   drop here

4. Evaluate with an unseen partner, without prior joint training.

   drop here

5. Pair policies that were not trained together and measure the result.

   drop here

CheckClear

2 · Calculate the Generalization Gap

The worksheet diagnostic from Partner Generalization: $\Delta_{\text{gen}} = J_{\text{familiar}} - J_{\text{unseen}}$.

2.1

Agent A scores $J_{\text{familiar}} = 92$ and $J_{\text{unseen}} = 68$. Calculate $\Delta_{\text{gen}}$.

AnswerCheck

2.2

Agent B scores $J_{\text{familiar}} = 86$ and $J_{\text{unseen}} = 79$. Calculate its gap.

AnswerCheck

3 · Calculate Cross-Play Performance

Percentage success. The diagonal is each policy with its own training partner.

	P1	P2	P3	P4
A	94	81	65	72
B	80	93	84	79
C	61	86	95	88
D	70	82	85	96

3.1

Calculate policy B’s mean over the three partners it did not train with.

Answer%Check

Need a hint?

3.2

Now the same for policy A.

Answer%Check

4 · Explore Familiar and Unseen-Partner Performance

Real results from the Adapt Lab: three agents, 40-step episodes, orders served. mostly-fetch is a held-out partner that behaves much like a training partner. alternate is a held-out partner that does not.

• Familiar mean
• mostly-fetch (near)
• alternate (far)

Agent

4.1

By how much does the Generalist beat the Specialist on familiar partners?

AnswerordersCheck

5 · Evaluate the Specialization Trade-off

A fixed-partner agent achieves 96% with its training partner and 54% with unseen partners.

A diverse-partner agent achieves 90% and 82% respectively.

5.1

Which would you deploy in a system where partners change frequently, and what trade-off are you accepting?

AnswersReveal

1

Answer to 1.

Partner diversity · agent modelling · ad hoc teamwork · zero-shot coordination · cross-play, matched to the descriptions in order.

2.1

Answer to 2.1.

$92 - 68 = 24$.

2.2

Answer to 2.2.

$86 - 79 = 7$.

3.1

Answer to 3.1.

$\tfrac{80 + 84 + 79}{3} = 81$.

3.2

Answer to 3.2.

$\tfrac{81 + 65 + 72}{3} = 72.7$.

4.1

Answer to 4.1.

$38.8 - 28.8 = 10$ orders.

5.1

Answer to 5.1.

No single right answer.

The diverse agent, in a system where partners change frequently. 82 against 54 on unseen partners is a 28-point difference, and unseen partners are the common case by assumption.

The trade-off: 6 points of peak performance with a familiar partner, 96 down to 90. You are giving up the ability to exploit one partner’s habits in exchange for competence across a range, which is the cost Training Partner Diversity named.

A strong answer also notes what the numbers omit: how the unseen partners were chosen. If they resemble the training partners, both figures are optimistic and the 82 in particular cannot be trusted.

Review with Flashcards

Revisit the 10 key concepts, equations, and intuitions from this chapter.

Open Adapt Flashcards