Accuracy-Latency Tradeoff in Approximate and Delayed Computing as a Game in Satisfaction Form

Approximate and delayed computing have emerged as promising paradigms to offer flexibility in computational accuracy and strategically differentiate tasks between edge and cloud execution to enhance resource utilization. However, these approaches introduce tradeoffs, potentially compromising accuracy on one hand and increasing latency on the other. In this paper, we explore the integration of approximate and delayed computing paradigms within the edge-cloud computing continuum. Users can either offload tasks for approximate computing at the edge or opt for exact but potentially delayed computing at the cloud. In this context, the joint problem of computation task offloading and data compression is formulated and solved as a non-cooperative Game in Satisfaction Form. Each user autonomously determines the amount of task to offload for either computing option and the percentage of data compression for approximate computing, aiming to achieve an acceptable accuracy-latency tradeoff. The formulated game admits a Satisfaction Equilibrium (SE) point, which is concluded using a Reinforcement Learning (RL)-based algorithm. Simulation results demonstrate the performance of the proposed task offloading framework in the achieved accuracylatency tradeoff compared against different offloading strategies.