Safe and Reliable Two-Stage Online Offloading Algorithm for Transportation Cyber-Physical Systems

In order to address the issue of insufficient task offloading decisions in vehicle networks of transportation cyber-physical systems (TCPS) because of multitasking and resource constraints, this study presents a quasi-Newton deep reinforcement learning-based two-stage online offloading (QNRLO) algorithm. Computer simulation experiments show that the approach performs exceptionally well in terms of convergence under various conditions and parameter configurations. Most of the trials are carried out in a simulated setting, and further real-world scenarios may be required to confirm the algorithm’s efficacy. This methodology initially implements batch normalization techniques to enhance the training process of the deep neural network, subsequently utilizing the quasi-Newton method for optimization to successfully approximate the ideal answer. According to the experimental results, the QNRLO algorithm’s loss function and normalized computation rate have converged after 2,000 iterations, demonstrating the algorithm’s excellent stability and dependability. The findings demonstrate that the computational load and training time can be further optimized by appropriately adjusting certain parameters without compromising convergence performance. Furthermore, the technique incorporates system transmission time allocation into the TCPS model, hence augmenting the model’s practicality. The proposed approach markedly enhances the efficiency and stability of job offloading compared to previous algorithms, effectively addressing task offloading challenges in TCPS and exhibiting considerable applicability and reliability.