MotiLearn: Contract-Based Incentive Mechanism for Heterogeneous Edge Collaborative Training

Abstract

Due to the computation capacity improvement of edge devices and the popularization of artificial intelligence, there has been a dramatic increase in topic of edge intelligence. In order to motivate edge users to participate in task release and collaborative training, it is crucial to design a feasible incentive mechanism so that the task publisher (Principal) can maximize his interests while guaranteeing the income of co-trainers (Agent). In this paper, we first characterize the computation model, agent's benefits, and principal's profits model by considering collaborative training features and user heterogeneity. Furthermore, through the asymmetry of information, we divide the edge intelligent environment into complete information and incomplete information scenarios, in which the principal only knows the distribution of the agents’ private information rather than the specific information. Subsequently by discussing the above two situations in a static environment, we obtain the basic laws of market operation and reveal the significance of introducing a dynamic environment. Then a two-period contract-based incentive mechanism (MotiLearn) is proposed to overcome the ratchet effects of long-term contracts under dynamic incomplete information environment. Finally, theoretical proof and numerical results show that the proposed incentive mechanism is feasible and can motivate both the principal and agents positively.