#rial

Question 1: Single choice

Foerster et al. (2016) proposed two ways for agents to learn a communication protocol, RIAL and DIAL. What does DIAL do that RIAL does not?

Each agent receives the average of the other agents' hidden states, so the number of agents may vary1
In centralised training it sends real-valued messages and passes gradients from receiver to sender2
It keeps exchanging real-valued messages and gradients between agents at execution time3
It picks each message as a discrete action and learns it by Q-learning inside each agent4