Abstract

How do we know if communication is emerging in a multi-agent system? The vastmajority of recent papers on emergent communication show that adding acommunication channel leads to an increase in reward or task success. This is auseful indicator, but provides only a coarse measure of the agent's learnedcommunication abilities. As we move towards more complex environments, itbecomes imperative to have a set of finer tools that allow qualitative andquantitative insights into the emergence of communication. This may beespecially useful to allow humans to monitor agents' behaviour, whether forfault detection, assessing performance, or even building trust. In this paper,we examine a few intuitive existing metrics for measuring communication, andshow that they can be misleading. Specifically, by training deep reinforcementlearning agents to play simple matrix games augmented with a communicationchannel, we find a scenario where agents appear to communicate (their messagesprovide information about their subsequent action), and yet the messages do notimpact the environment or other agent in any way. We explain this phenomenonusing ablation studies and by visualizing the representations of the learnedpolicies. We also survey some commonly used metrics for measuring emergentcommunication, and provide recommendations as to when these metrics should beused.

 

Introduction (beta)

None

 

Conclusion (beta)

None