My question is regarding the paper Learning to Communicate with Deep Multi-Agent Reinforcement Learning. Can anyone explain what is the significance of Colour-digit MNIST game in the paper? I understand from here that the agent has to choose an action that represents
if (case 1) color of agent 1 == parity of agent 2
or (case 2) color of agent 2 == parity of agent 1.
The reward for case 1 is twice that of case 2. Hence, it optimal case, both agents should learn (case 1) to communicate their parity to other agent via the message and learn to check the received message against its own color. But if this seems to mean that case 2 is just acting like a noise or additional challenge in attaining optimal solution. Even if the case 2 terms did not exist, still the agents had a learn non-trivial communication protocol. Having said that, does case 2 term has any fundamental implications/need?
