Multi-criteria deep Q-Network scheduling for load balancing in heterogeneous edge computing environments
2026, vol.18 , no.4, pp. 11-22
Article [2026-04-02]
Task scheduling in heterogeneous edge computing is an online, NP-hard problem that must jointly account for server hardware heterogeneity and time varying load. Deep reinforcement learning schedulers learn server capability only from experience, while multi criteria decision making yields informative but static rankings. This paper proposes MCDRL-HLB, which augments the state of a Double Deep Q-Network agent with Complex Proportional Assessment utility scores derived from server hardware. Four conditions, DQN, DDQN, COPRAS-DQN and MCDRL-HLB, were evaluated under identical settings in EdgeCloudSim across three random seeds. MCDRL-HLB lowered mean task failure rate from 22.67% to 18.42% against the DQN baseline and cut cross-seed variance by a factor of 1.8, improving both accuracy and convergence consistency. Generalisation outside the training load remained seed sensitive, a boundary reported openly alongside the demonstrated gains.
edge computing, deep reinforcement learning, task scheduling, multi-criteria decision-making, load balancing
https://doi.org/10.59035/RHQJ9694
Muhammad Rehan Rasheed, Rabia Noor Enam, Rukaiya Rukaiya, Muhammad Naseem. Multi-criteria deep Q-Network scheduling for load balancing in heterogeneous edge computing environments . International Journal on Information Technologies and Security, vol.18 , no.4, 2026, pp. 11-22. https://doi.org/10.59035/RHQJ9694