Learning Communication Policies for Knowledge Transfer between Agents
Abstract
We present an agent model in the predictive coding framework that selectively communicates with other agents to predictthe state of its environment efficiently. Selective communication is a challenge when the internal models of other agentsare unknown and unobservable. Communication helps agents to transfer the knowledge they have acquired in differentsituations. Recognition of daily activities of individuals living in different homes served as a testbed for evaluating themodel. Two publicly-available datasets, collected from unique homes, are used. Behavioral patterns of individuals in thosehomes are also unique. Each home is assumed to be monitored by an agent. We experimentally show that the agents cantransfer knowledge by communicating the most informative messages. The messages are interpretable. The agents learnpatterns of daily activities for any individual, and communicate using a vocabulary of words. Our model is more accuratethan traditional transfer learning models for the same task.