#maddpg

Question 1: Single choice

MADDPG trains $N$ agents with centralised training and decentralised execution. What does agent $i$ 's critic receive, and what is used once training is over?

The critic sees all agents' observations and actions; afterwards each actor uses only its own observation1
One critic shared by all agents sees the global state, so MADDPG needs a common team reward2
The critic sees only agent $i$ 's observation and action, and the actors share observations when acting3
Critic and actor both see every agent's observations and actions, in training and in execution4

Timed test