Energy-Efficient Intelligence Sharing in Intelligence Networking-Empowered Edge Computing: A Deep Reinforcement Learning Approach

Advanced artificial intelligence (AI) and multi-access edge computing (MEC) technologies facilitate the development of edge intelligence, enabling the intelligence learned from remote cloud to network edge. To achieve automatic decision-making, the training efficiency and accuracy of AI models are crucial for edge intelligence. However, the collected data volume of each network edge node is limited, which may cause the over-fitting of AI models. To improve the training efficiency and accuracy of AI models for edge intelligence, intelligence networking-empowered edge computing (INEEC) is a promising solution, which enables each network edge node to improve its AI models quickly and economically with the help of other network edge nodes' sharing of their learned intelligence. Sharing intelligence among network edge nodes efficiently is essential for INEEC. Thus in this paper, we study the intelligence sharing scheme, which aims to maximize the system energy efficiency while ensuring the latency tolerance via jointly optimizing intelligence requesting strategy, transmission power control and computation resource allocation. The system energy efficiency is defined as the ratio of model performance to energy consumption. Taking into account the dynamic characteristics of edge network conditions, the intelligence sharing problem is modeled as a Markov decision process (MDP). Subsequently, a twin delayed deep deterministic policy gradient (TD3)-based algorithm is designed to automatically make the optimal decisions. Finally, by extensive simulation experiments, it is shown that: 1) compared with DDPG and DQN, the proposed algorithm has a better convergence performance; 2) jointly optimizing intelligence requesting strategy, transmission power control and computation resource allocation helps to improve intelligence sharing efficiency; 3) under different parameter settings, the proposed algorithm achieves better results than the benchmark algorithms.