Meta-Learning-Based Deep Learning Model Deployment Scheme for Edge Caching

Abstract

Recently, with big data and high computing power, deep learning models have achieved high accuracy in prediction problems. However, the challenging issues of utilizing deep learning into the content’s popularity prediction remains open. The first issue is how to pick the best-suited neural network architecture among the numerous types of deep learning architectures (e.g., Feed-forward Neural Networks, Recurrent Neural Networks, etc.). The second issue is how to optimize the hyperparameters (e.g., number of hidden layers, neurons, etc.) of the chosen neural network. Therefore, we propose the reinforcement (Q-Learning) meta-learning based deep learning model deployment scheme to construct the best-suited model for predicting content’s popularity autonomously. Also, we added the feedback mechanism to update the Q-Table whenever the base station calibrates the model to find out more appropriate prediction model. The experiment results show that the proposed scheme outperforms existing algorithms in many key performance indicators, especially in content hit probability and access delay.