Energy-Efficient Cooperative Inference in Multidevice Edge Networks: A Lyapunov-Based Approach

In mobile edge computing (MEC) networks, the limited computational capabilities of edge devices present challenges in achieving energy-efficient and high-accuracy inference, highlighting the critical role of cooperative inference with an edge server. This study proposes a cooperative inference framework in which edge devices employ dual confidence thresholds to filter ambiguous images, which are stored in task buffer queues and offloaded to the edge server for more accurate inferences. We conduct a numerical analysis of the inference accuracy and energy consumption for the proposed cooperative inference method and formulate a joint optimization problem to determine the optimal confidence thresholds, offloading ratio, transmit power, and transmission/reception time, aiming to minimize energy consumption while ensuring accuracy and queue stability. We employ a Lyapunov-based optimization approach to convert this optimization problem into a time-independent real-time decision problem. Subsequently, we decompose it into tractable subproblems and propose an iterative algorithm based on the block coordinate descent method to derive suboptimal variables efficiently. The experimental results reveal a tradeoff between inference accuracy and cooperation penalty depending on the confidence thresholds. In particular, optimizing these thresholds in conjunction with radio resources significantly reduces energy consumption and queue backlog while maintaining the required accuracy in diverse MEC environments, compared with the conventional inference methods.