Multi-band provisioning in dynamic elastic optical networks: a comparative study of a heuristic and a deep reinforcement learning approach

The blocking performance of a heuristic and a deep reinforcement learning approach for resource provisioning in a dynamic multi-band elastic optical network is evaluated. The heuristic is based on a previous proposal that prioritises the use of band C, then L, S, and E, in that order. The deep reinforcement learning approach uses a deep Q-network (DQN) agent trained on different multi-band scenarios. Results show, as expected, a significant decrease in blocking probability when moving from the C-band only scenario to the multi-band scenarios (C+L, C+L+S, C+L+S+E). However, the DQN agent did not outperform the heuristic. The lower performance of the agent, also observed in some previous works in optical networks, highlights the need for further research on how to better configure agents and improve the network representation used by the optical network environments