Semantic Task Offloading for IoT in Heterogeneous Meta-Computing Environments: A DRL Approach with LLM-Enhanced Contextual Awareness

Abstract

The continuous increase in Internet of Things (IoT) devices in heterogeneous meta-computing environments has led to increased interest in more efficient task offloading methods to mitigate the effects of resource constraints in processing, memory, and battery life. Existing offloading methods often lack semantical offloading decisions and do not consider the heterogeneous and dynamic nature of these environments, leading to reduced results. This paper presents a semantic representation-enhanced DRL framework that integrates Deep Reinforcement Learning (DRL) and Large Language Models (LLMs) to enable more efficient task offloading. An LLM acts as a semantic encoder, converting raw task descriptions into embedding vectors that capture task intent, dependencies, and computational requirements. Based on this description, a DRL-based agent makes the best offloading decision that can be made given the current state of the system. This paper reports a 37.2% reduction in latency and a 23.8% reduction in energy consumption compared to DQN, along with a 17.4% improvement in task success rate over QMIX under high-load conditions.