Energy Efficient Training Task Assignment Scheme for Mobile Distributed Deep Learning Scenario Using DQN

Recent years, applications of virtual or augmented reality, image or speech recognition which depend on deep learning are popular in mobile devices. The accuracy and training time are further affected by two factors: hardware performance and the dataset. The hardware resource of individual devices including CPU, GPU and memory, can't meet the demand of deep-learning training itself. However, it is possible to assign the training tasks to multiple devices. Besides, the best data of training are on the phone we used because the behavior of model is relies on the correlation of dataset. In this paper, we propose a deep Q-Learning based training task assignment scheme for mobile deep learning scenario with the objective of minimizing energy consumption under the quality of service (QoS) constraint in the dynamic environment. Particularly, an evaluation mechanism is presented to help participant selection and a deep reinforcement learning framework is used for resource allocation due to high-dimensional state space control problem in dynamic environments. Finally, the simulation results compared with different schemes prove the feasibility of the proposed framework.