.png){kind=icon}
.png){kind=icon}
Return
EdgePlus: A Multiagent Reinforcement Learning Framework for Dynamic Task Allocation in 6G Edge Computing
.png){kind=icon}
PRE
AI
DOI：10.1109/JIOT.2026.3656811.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In smart-city air-quality monitoring, millions of IoT sensors generate time-sensitive data that must be processed within strict end-to-end deadlines, under dynamic mobility conditions, and with strong security guarantees. In this study, we present EdgePlus, a unified sixth-generation (6G) framework that couples 1) masked cooperative PPO (CCPPO) enforcing feasibility at sampling time (routing order, slice quotas, resource limits), 2) a constricted particle swarm optimization (PSO) outer loop that escapes local optima under strict SLA gates, 3) a lightweight advisory layer (Nash sharing, Ziegler–Nichols-initialized PID, and leader election) that biases logits without breaking masks, and 4) a two-tier blockchain-lite ledger [edge proof-of-authority (PoA), cloud Byzantine fault-tolerant BFT)] whose reputation and debt signals feedback into state, reward, and masks. In a 100 km2 city-scale simulator (ten clusters; three load regimes of 50 000, 100 000, and 150 000 tasks). Under light load, all methods reach 100% success with seconds-level latencies and no observable penalty from masking or security. At stress (150k), EdgePlus achieves the highest on-time completion rate (82.77%), highest throughput (206.93), lowest violation rate (17.23%), and lowest 15 percentile latency ( $L_{\!p50}$ of 53.23 s). Ablations confirmed that feasibility masks dominate performance: removing them increases violations by 5.5%–10.0% points and tail latency ( $L_{\!p95}$ ) by 26–47 s, while advisory and PSO deliver smaller but consistent gains. The PID controller remains robust under ±20% gain drift (settling $\approx 2.7$ –2.9 s) by design-time Jury certification. In security experiments, baseline confidentiality, integrity, and availability (CIA)/auth attacks are mitigated at 98%–99% and advanced (nonphysical-layer) threats achieve 70%–83% defense with strong efficiency. Edge-tier PoA validation adds a median $\approx 0.39$ ms per partial block for validator sets up to 10 and remains sub-ms with partitioned sub-quorums up to 30, while cloud BFT provides accountable finality off the critical path. In short, EdgePlus delivers a single, coherent scheduling–control–security stack for 6G edge computing, built on a feasibility-first design that preserves reliability, locality, and security at scale.
Keywords:
Blockchain-lite security
constrained reinforcement learning
low-latency networks
multiagent systems
resource allocation
sixth-generation (6G) edge computing
Journal
IF:
8.9
Papers:
1.4W
Citations:
7.8W