Efficient task offloading in satellite edge computing requires coordinated allocation of communication and computing resources. Search methods based on large language models (LLMs) have been shown to generate high-quality resource allocations, but the latency of individual model requests and repeated model interactions can lead to substantial solver overhead. We propose budgeted joint-action search (BJAS), an integrated decision-model-guided solver that jointly optimizes satellite offloading and power allocation to minimize average task-completion latency under explicit search budgets. Its bounded candidate pools combine single-terminal updates with power-inheriting offload replacements for batched selection, retaining two-terminal coordination without enumerating additional power vectors. Across 60 completed runs per method under the workflows' respective configurations, BJAS reduces mean total solver time by 97.16% relative to LLM-generated offloading while achieving a mean task-completion latency close to that of the LLM baseline. BJAS also achieves lower mean task-completion latency than the tested genetic algorithm and deep deterministic policy gradient baselines.
翻译:暂无翻译