Multi-Orchestrator Mobile Edge Learning: Designing Energy-Efficient Task Allocation and Incentive Schemes

Abstract

Mobile Edge Learning (MEL) is a decentralized collaborative learning paradigm that features distributed training of Machine Learning (ML) models over resource-constrained edge devices (e.g., Internet of Things (IoT) devices). MEL enables such devices to either learn a shared model without sharing data, or distribute the learning model along with the data to other IoT devices and utilize their available resources. In the former case, IoT devices (aka learners) need to be assigned an orchestrator to facilitate decentralized learning and models' aggregation from different learners. Whereas in the latter case, IoT devices act as orchestrators and look for learners with available resources to distribute the learning task and utilize their resources.

However, in MEL, the coexistence of multiple learning tasks with different datasets may arise, which is referred to as multi-orchestrator MEL. The heterogeneity in edge devices' capabilities will require the joint optimization of the learners-orchestrators association and task allocation. Moreover, the performance of each learning task deteriorates without the availability of sufficient training data or computing resources. Therefore, it is crucial to motivate the edge devices to become learners and offer their computing resources, and either offer their private data or receive the needed data from the orchestrator and participate in the training process of a learning task.

To this end, we aim to develop an energy-efficient framework for orchestrators-learners association and learning task allocation, in which each orchestrator gets associated with a group of learners with the same learning task based on their communication channel qualities and computational resources, and allocate the tasks accordingly. Afterward, we propose an incentive mechanism, where we formulate the orchestrators-learners interactions as a two-round Stackelberg game to motivate the participation of the learners. In the first round, the learners decide which learning task to get engaged in, and then in the second round, the training parameters and the amount of data for training are decided in case of participation such that their utility is maximized. Finally, numerical experiments have been conducted to evaluate the performance of the proposed energy-efficient framework and the proposed incentive mechanism.