Deep reinforcement learning-based multi-objective routing and resource allocation strategy in quantum key distribution optical networks

Abstract

Quantum key distribution (QKD) optical networks, relying on the information-theoretic security guaranteed by the fundamental laws of quantum physics, have become a core solution for ensuring data transmission security in critical fields such as finance and energy. However, QKD optical networks have inherent characteristics like severe resource constraints, dynamic changes in network environments. The collaborative optimization of routing, wavelength, and time-slot assignment (RWTA) has thus become a key bottleneck restricting their practical deployment. Traditional algorithms suffer from high complexity and poor adaptability in large-scale dynamic networks, while existing deep reinforcement learning (DRL) schemes insufficiently consider network security states, failing multi-objective co-optimization. Therefore, this paper proposes a multi-objective secure proximal policy optimization algorithm (MOS-PPO) based on DRL to solve the routing, wavelength, and time slots assignment problem in QKD optical networks. The MOS-PPO algorithm expands the state space by introducing a network security state vector that includes quantum bit error rate, risk coefficient, node capacity, and processing capability. It designs a multi-objective reward function, and establishes a security constraint mechanism to achieve the collaborative optimization of blocking probability (BP), resource utilization (RU), and security performance. Simulations on NSFNET and UBN24 topologies show MOS-PPO outperforms DRL-PPO algorithm, Security-aware Load balancing, first fit algorithm, and random fit algorithm. On NSFNET, average BP drops by 20.4%, 22.1%,28.4%, 36.7% and RU rises by 4.0%,3.9%, 22.2%, 31.0%, on UBN24, BP decreases by 9.3%, 11.4%, 18.8%, and 27.4% and RU increases by 3.8%, 21.2%, 10.3%, and 29.5%, with higher stable security margins. This study verifies the high adaptability and multi-objective optimization ability of the MOS-PPO algorithm in dynamic network environments, providing an efficient RWTA solution for the practical deployment of QKD optical networks.