Offloading Budget-constrained Task to Edge via Reinforced Embedding

In edge computing applications, partitioning user tasks into subtasks and offloading them to multiple edge servers for parallel execution is a promising strategy to reduce latency. Despite its potential, existing research often overlooks the challenges introduced by budget constraints and heterogeneous processing capabilities of edge servers. This paper proposes the Partition-Allocation-Selection (PAS) problem, which makes decisions on task partition, subtask allocation, and selection of executors under budget constraints to minimize task makespan. This problem is not only NP-hard but also characterized by intricate interdependencies among these decisions. To address computational challenges, we propose ReMeNet, a novel framework based on reinforcement learning and representation learning, which reformulates PAS as a learning problem within an embedding space. ReMeNet introduces two innovations: (1) it unifies task partitioning, subtask allocation, and executor selection into a single-step decision-making process using continuous-action reinforcement learning, which efficiently encodes their interdependencies, ensuring disjoint task partitions, and managing the high dimensionality of the decision space, and (2) it adopts a representation learning strategy to produce supervision-free, refined and interpretable action representations compared to traditional discrete-action methods. Extensive numerical experiments demonstrate that ReMeNet significantly outperforms state-of-the-art approaches, underscoring its potential to advance intelligent task offloading in edge computing.