Deep Neural Network-Based Reliability Estimation and Optimization of Networks

Summary & ConclusionsModern networks operating in dynamic environments face significant challenges in ensuring reliable performance under cost and design constraints. In all-terminal networks, where complete connectivity between all nodes must be maintained, reliability estimation becomes increasingly complex as network size grows. This work presents a data-driven framework that integrates Deep Neural Network (DNN) and Simulated Annealing (SA) to efficiently estimate and optimize all-terminal network reliability. A DNN is trained to approximate reliability from network topology and link-reliability inputs. The training dataset is generated using Monte Carlo (MC) simulation across mesh, ring, and star ten-node topologies over a range of link reliability values, enabling rapid evaluation of configurations without repeated simulation during optimization. Under five-fold cross-validation (CV), the DNN achieved RMSE of 0.018569, and the resulting estimates demonstrate clear topology sensitivity.For design, the trained DNN is embedded within SA to explore feasible configurations and identify topologies that maximize reliability while satisfying a cost constraint. Across ten independent SA runs, the best-of-run reliability values ranged from 0.998563 to 0.999663, with the global best design achieving a reliability of 0.999663 using 11 links at a total cost of 1485 units. Within the feasible design space (1485–1890 cost units), reliability varied by no more than 0.11%, indicating stable convergence and consistent optimization performance. This approach provides a scalable and practical solution for optimizing reliability-aware network design, combining fast estimation with efficient search to support informed decision-making in complex network environments.