Title:Winning an Election: On Emergent Strategic Communication in Multi-Agent Networks

Abstract:Humans use language to collectively execute abstract strategies besides using it as a referential tool for identifying physical entities. Recently, multiple attempts at replicating the process of emergence of language in artificial agents have been made. While existing approaches study emergent languages as referential tools, in this paper, we study their role in discovering and implementing strategies. We formulate the problem using a voting game where two candidate agents contest in an election with the goal of convincing population members (other agents), that are connected to each other via an underlying network, to vote for them. To achieve this goal, agents are only allowed to exchange messages in the form of sequences of discrete symbols to spread their propaganda. We use neural networks with Gumbel-Softmax relaxation for sampling categorical random variables to parameterize the policies followed by all agents. Using our proposed framework, we provide concrete answers to the following questions: (i) Do the agents learn to communicate in a meaningful way and does the emergent communication play a role in deciding the winner? (ii) Does the system evolve as expected under various reward structures? (iii) How is the emergent language affected by the community structure in the network? To the best of our knowledge, we are the first to explore emergence of communication for discovering and implementing strategies in a setting where agents communicate over a network.