FC-DDQN for Multi-User Computation Offloading and Resource Allocation in Mobile Edge Computing

Abstract

This repository contains the software and reproducibility materials associated with the study “Factorized Coordinated Double Deep Q-Learning for Multi-User Computation Offloading and Resource Allocation in Mobile Edge Computing.” The proposed Factorized Coordinated Double Deep Q-Network (FC-DDQN) separates the binary local/offloading decision from the discrete MEC CPU-resource decision using independently parameterized value networks. A lightweight MEC-side coordinator ranks simultaneous offloading proposals according to their learned Q-value advantage and converts them into a resource-feasible joint allocation. Coordination is applied during training and evaluation so that the actions stored in replay memory correspond to those executed in the MEC environment. The repository includes the MEC simulation environment, the FC-DDQN implementation, conventional DQN and DDQN components, coordination and ablation scripts, experimental configurations, the original summary results from ten independent FC-DDQN training runs, validation-selected checkpoints, and the final publication figures. Under the nominal configuration, FC-DDQN achieves a mean normalized constraint-aware system objective of approximately 0.02005 across ten independent training runs, with zero mean MEC overload. The repository also documents the validation and held-out test seeds used in the experiments. The final Figures 4–8 are preserved as used in the revised manuscript and are not automatically regenerated by this repository. Some original trajectory-level outputs for conventional baselines and intermediate ablation configurations were not retained and are therefore not represented as available raw data.