Context-Aware Effective Communications

We investigate the effective communications problem, in which the goal of the transmitter is to impact the actions of the receiver over a time frame in order to maximize the prescribed reward function. This problem is formulated as a multi-agent partially-observable Markov decision process (MA-POMDP), where one agent can communicate to the other through a noisy communication channel. In this paper, we show that not only should the communication scheme be jointly designed with the underlying learning objective, but the context of the problem can also be exploited to achieve even greater effectiveness. Here, the context refers to a function of the state of the environment that is available to both agents. We then show that, using different communication schemes depending on the context is beneficial to the effectiveness of the solution. We emphasize that this is different from sending different messages at different states; with the proposed context-aware communication scheme, the same message is interpreted differently by the receiver depending on the context, similarly to human communications, where the meaning of a sentence may change depending on the context.