#cooperative

Question 1: Select all that apply

In a fully cooperative task all agents receive one shared team reward. VDN and QMIX learn a joint value $Q_{tot}$ from per-agent utilities $Q_a(\tau^a, u^a)$ , where $\tau^a$ is agent $a$ 's own action-observation history. Which statements are true? Select all that apply.

QMIX can represent any joint value, even one where an agent's best action depends on the others' actions1
VDN sets $Q_{tot}$ to the sum of the agents' $Q_a$ 2
QMIX requires $\partial Q_{tot} / \partial Q_a \ge 0$ , so the agents' greedy actions together maximise $Q_{tot}$ 3
QMIX keeps the mixing weights non-negative, using hypernetworks that take the state $s$ 4
At execution each agent needs the global state, since the mixing network uses it5

Timed test