A reinforcement learning multiagent architecture prototype for smart homes (IoT)

Abstract

Continuous technology progress is fueling the delivery of new and less expensive IoT components, providing a variety of options for the Smart Home. Although most of the components can be easily integrated, achieving an optimal configuration that prioritizes environmental goals over individual performance strategies is a complex task that requires manual fine tuning. The objective of this work is to propose an architecture model that integrates reinforcement learning capabilities in a Smart Home environment. In order to ensure the completeness of the solution, a set of architecture requirements was elicited. The proposed architecture is extended from the IoT Architecture Reference Model (ARM), with specific components designed to coordinate the learning effort, as well as data governance and general orchestration. Besides confirming the fulfillment of the architecture requirements, a simulation tool was developed to test the learning capabilities of a system instantiated from the proposed architecture. After six million and four hundred thousand execution cycles, it was verified that system was able to learn in most of the configurations. Unexpectedly, results show very similar performance for collaborative and competitive environments, suggesting that a more varied selection of agent scenarios should be tested as an extension of this work, to confirm or contest Q-Learning hypothesis.