Joint Service Function Chain Embedding and Routing in Cloud-based NFV: A Deep Q-Learning Based Approach

By interacting with the environment through trial and error, a deep reinforcement learning (DRL) -based controller is able to learn the patterns underlying a highly dynamic environment, paving the way for self-optimized decision making. Leveraging this capability, we propose a Deep Q-learning (DQL)-based algorithm to perform joint service function chain (SFC) integration and routing tasks for service requests, each with specific stringent end-to-end delay and bandwidth constraints in a cloud-based distributed Network Function Virtualization (NFV). Our design goal is to maximize the number of dynamically provisioned service requests in the network over a lime horizon, thus indirectly contributing to maximizing network throughput. Numerical results show that the trained DQL-based algorithm achieves over 95% of the average request acceptance rate and over 96% of the offered load. This performance result is comparable to that of a Resource-Constrained Shortest Path algorithm while achieving a much shorter execution time, approximately 10 times faster.