TY  - EJOU
AU  - Ma, Zengqi 
AU  - Zhao, Wendong 

TI  - Value-of-Information Driven Distributed Offloading Decision for UAVs: A Behavior Prediction-Based Deep Reinforcement Learning Approach
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - In large-scale monitoring scenarios with multiple unmanned aerial vehicles (UAVs) and offloading nodes, concurrent offloading of collected data to the same node within a short time window will cause excessive node load, prolonged processing delay, and degraded information timeliness. To address this issue, this paper considers the impact of users’ personalized demands on information timeliness and introduces the concept of Value of Information (VoI). A user demand-oriented dynamic VoI model is established to accurately describe and evaluate information timeliness. The multi-UAV offloading node selection problem is formulated as a semi-Markov decision process (SMDP), and a Behavior Prediction-based Distributed Offloading Decision (BP-DOD) algorithm is proposed. By predicting UAV offloading behaviors and constructing conflict costs, the algorithm explicitly characterizes the competition for computing resources caused by the mismatch between the generation order and execution order of different UAV decisions, thereby significantly improving convergence stability. Furthermore, Dueling DQN, Double DQN and reward shaping techniques are integrated to further improve algorithm performance. Simulation results show that compared with benchmark algorithms, the proposed algorithm can increase the average VoI by up to 0.777 and reduce the information invalidation rate by up to 86.6 percentage points, and exhibits favorable adaptability across different scales of UAVs and edge nodes.
KW  - Unmanned aerial vehicle; value of information; multi-access edge computing; user demand; deep reinforcement learning

DO  - 10.32604/cmc.2026.088063