Deep reinforcement learning with dual-Q and Kolmogorov–Arnold Networks for computation offloading in Industrial IoT

Abstract

In the industrial internet of things, the rapid development of smart mobile devices and 5G network technology has driven the application of mobile edge computing, reducing the delay in task computation offloading to some extent. However, the increasing complexity of the IIoT environment presents challenges for communication management and offloading performance. To achieve efficient computation offloading communication, we designed a cloud–edge-device IIoT system model, utilizing Voronoi diagrams to partition the service areas of edge servers, thereby adapting to the complex IIoT environment and improving communication efficiency. Additionally, considering that different offloading strategies may result in varying levels of offloading security risks, we developed a principal component analysis-based offloading security evaluation model (PCA-OSEM) to analyze potential security risks during the offloading process and identify key factors. Finally, to optimize offloading strategies to reduce offloading delay and security risks, we proposed a dual-Q with Kolmogorov–Arnold networks in deep reinforcement learning computation offloading (D2KCO). This method enhances the neural network’s approximation capability and training stability. Experimental results show that the proposed PCA-OSEM is effective, and the D2KCO method can reduce offloading delay by 13% and 23.52% compared to the D3PG and DDPG algorithms, respectively, while also reducing security risks.