An MARL-Based Distributed Learning Scheme for Capturing User Preferences in a Smart Environment

Providing a personalized service to a user in a smart environment has been one of the key goals in the area of pervasive computing. The proliferation of individually developed smart devices in the name of Internet of Things opens up a possibility of providing personalized services to a user in an autonomous and distributed manner. As a user's task often involves services supported by multiple devices, capturing a device-specific service preference is not enough to maximize a user's comfort. In this paper, we propose a distributed learning scheme for capturing multiple device service preferences in a smart environment. We exploit multi-agent reinforcement learning (MARL) method where each smart device acts as a reinforcement learning agent to incrementally and cooperatively capture a user specific preference of a task. Experiments confirm that smart devices with the proposed scheme are able to capture multiple device service preferences from a small number of interactions with a user and an environment. Also, the proposed transfer learning method improves learning performance for a new task.