|
چکیده
|
Recently, the advancement of networking technology has moved on to flexible, programmable, and intelligent architectures. Of these emerging networking architectures, Software Defined Networks (SDNs) have emerged as an innovative paradigm that separates the network control plane from the data plane. This makes centralized control possible and network setup easier and more flexible. However, although SDNs offer unparalleled control and programmatic functionalities, challenges also abound with regards to routing in these networks, especially in environments with dynamic and unpredictably changing traffic patterns. Conventional routing strategies that were typically developed for statically or semi-statically configured networks are not very efficient with sudden congestion patterns, changing traffic patterns, and network failures.
Deep Reinforcement Learning (DRL), on the other hand, has been recognized as an efficient way for overcoming the mentioned issues. DRL models learn from the interactions with the network and rewards/penalties given by the network and can learn and discover autonomously the most efficient routing methods that adapt dynamically to the evolving network conditions on their own. In fact, DRL models for routing can consider complicated and time-varying traffic patterns in the network in contrast to traditional methods for optimization problems and algorithms. Finally, as mentioned above, among other problems associated with reinforcement learning models, DRL models for routing have the significant problem of not being transparent enough for network operators and understanding their decisions or actions.
This research aims at filling this gap within the context of explainable deep reinforcement learning for SDN routing. The major objective of this research is the formulation of a model that not only aims at the optimization of routing efficiency but also at the interpretation of the decision-making process of the deep reinforcement learning agent. I
|