A Multi-Objective Multi-Agent RL-Based Approach for Task Offloading in IoT-Enabled Classrooms

With the widespread deployment of smart sensors and edge computing, Internet of Things (IoT) has become an im-portant enabler to facilitate many conventional systems, such as e-learning, e-health, and assisted driving. Among these systems, one of the most impactful, IoT-enabled practices is e-learning systems, which aim to capture real-time behaviors of students via a multitude of sensors in accordance with course materials, and to further provide adaptive teaching effectiveness. However, e-learning system, specifically an IoT-enabled classroom, faces three important challenges: (1) how to capture the real-time sensing information; (2) how to collect and differentiate jobs with delay constraints, such as face recognition, multi-modal attention detection, etc; and (3) how to allocate computing resources of edge nodes to satisfy the delay requirements of jobs that consist of multiple tasks. To tackle the first two challenges, we design and implement a novel IoT-enabled, e-learning system, named Yo-Yo, with edge computing to collect real-time jobs and tasks for further data analysis. In Yo-Yo system, both job completion rate and energy consumption are considered as quality-of-service (QoS) metrics. Therefore, for the third challenge, we propose a new multi-objective, multi-agent reinforcement learning method to jointly optimize these two objectives. Experiments results indicate that for each device, around 4,200 jobs are generated, with varying tasks associated with each job. The average job completion rate can achieve up to 94% over 10 devices; for dealing with complex task dependency, our approach can improve by 24 %, while reducing the ineffective energy consumption by 70 % against to single-agent based methods.