NPTSN: RL-Based Network Planning with Guaranteed Reliability for In-Vehicle TSSDN

To achieve strict reliability goals with lower redundancy cost, Time-Sensitive Software-Defined Networking (TSSDN) enables run-time recovery for future in-vehicle networks. While the recovery mechanisms rely on network planning to establish reliability guarantees, existing network planning solutions are not suitable for TSSDN due to its domain-specific scheduling and reliability concerns. The sparse solution space and expensive reliability verification further complicate the problem. We propose NPTSN, a TSSDN planning solution based on deep Reinforcement Learning (RL). It represents the domain-specific concerns with the RL environment and constructs solutions with an intelligent network generator. The network generator iteratively proposes TSSDN solutions based on a failure analysis and trains a decision-making neural network using a modified actor-critic algorithm. Extensive performance evaluations show that NPTSN guarantees reliability for more test cases and shortens the decision trajectory compared to state-of-the-art solutions. It reduces the network cost by up to 6.8x in the performed experiments.