This survey delivers the first systematic exploration of how artificial intelligence can bolster SR, from traffic classification and segment-list computation to fast reroute, service-function chaining, and multi-domain orchestration by spanning supervised and unsupervised learning, reinforcement learning, and hybrid pipelines that fuse forecasting, neural optimization, and heuristic search.
Adaptive routing techniques that go beyond the constraints of conventional algorithms are required due to the extraordinary increase in network traffic. The advancements in deep reinforcement learning (DRL) techniques enforce the dynamic learning of the paths in network. This paper presents a hybrid DRL based Graph Neural Network combined with Soft Actor Critic(GNN-SAC) method in a dynamic Software Defined Network(SDN) environment using M/M/1 queuing model. Its performance is compared with other four DRL techniques-Advantage Actor-Critic (A2C), Deep Deterministic Policy Gradient (DDPG), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC). The experimental results demonstrate that the hybrid GNN-SAC achieves the best path cost of 4, highest throughput of 25.18 Mbps, and lowest packet loss of 6.02\%, whereas SAC achieves the lowest delay of 14.74 ms and jitter of 1.16 ms, however, A2C achieves the fastest training time of $0.23 s$ and best link utilization of 45.58\%. This platform also provides real-time animated packet routing visualization with live failure simulation capabilities.
Ramani Kasarapu, N. K. Kumar, A.G.Srideep Reddy et al.· International Conference on...· 0 citations
A QoE-aware framework for Multi-Access Edge Computing-enabled Open Radio Access Network (O-RAN) architectures, combining a graph attention network (GAT) encoder, distributed multi-agent DRL, and privacy-preserving FL, while transitioning control from Quality of Service (QoS) to QoE metrics is proposed.
Manoj Prasad Kunasegran, Wai Leong Pang, S. K. Phang· IEEE Access· 0 citations
Networks with highly dynamic data transmission demands and network topologies are common in real world. A fundamental problem in such networks is achieving scalable traffic allocation to maximize long-term total throughput under link capacity constraints. However, state-of-the-art (SOTA) works lack scalability. This is primarily due to two reasons in large-scale networks: first, they require solving constrained optimization problems online, which leads to high decision latency; second, they rely on reinforcement learning algorithms for policy optimization, which are inefficient in exploration and challenging to train effectively. To address these issues, we propose the Fast Networked Control (FNC) policy framework, which firstly utilizes parallelizable neural network modules to process the state and generate raw decisions, followed by basic operations such as normalizations and comparisons, which do not require iteration or optimization, to obtain decisions that satisfy the constraints. Hence, FNC policy avoids solving constrained optimization problems and supports parallel execution, significantly reducing decision latency. Furthermore, this policy preserves gradient flow and supports backpropagation, which enable us to design an imitation learning algorithm to efficiently train the policy in an end-to-end manner. Experiments in large-scale networks show that our FNC policy achieves an average 8% improvement in demands satisfaction and 10 times reduction in decision latency versus SOTA works.
Zhaoxing Yang, Guiyun Fan, Anjie Cao et al.· IEEE Transactions on Network...· 0 citations
This work proposes Double-Channel Graph Attention (DCGA), an end-to-end reinforcement learning framework that isolates network reachability and demand-service logic into separate graph channels and constructs valid routes using a simulator-coupled, constraint-informed decoder.
As the modern network infrastructure continue to become more and more complex, massive set of firewall rules have been produced, which causes large package classification latency and poor resource utilization. Firewall rule ordering is an NP-hard mathematical programming problem in which the wrong sequencing of firewall rules maximizes the cost of matching, as well as presents the possibility of policy violation. The current meta-heuristic algorithms are said to be constrained by predetermined traffic patterns and unresponsive to real time. In this paper, an adaptive Reinforcement Learning (RL) framework is introduced to the dynamic optimization of firewall rules. The issue is presented as a Markov Decision Process (MDP) and Deep Q-Network (DQN) agent can observe the frequency of rule hits and dependency restrictions and propose the best reorganization. In order to solve the scalability issue, A state-space dimensionality reduction algorithm is adopted, where we target high-traffic rule sets. The Directed Acyclic Graph (DAG) is used to make sure that the reordering of rules does not compromise the semantics of security policies. The experimental results prove that the throughput is increased and the cost of rule comparisons are decreased as compared to fixed configurations. Our framework is better in changing traffic dynamics.
Jonnalagadda Syam Babu, K. V. Shilpa, M.Hanish Kumar et al.· 2026 7th International Confe...· 0 citations
Future sixth-generation (6G)-oriented networks require programmable control that can adapt routing to latency and congestion without unsafe online exploration. This study evaluates offline multi-agent deep deterministic policy gradient (MADDPG) with behavior-adjusted training rewards for latency-aware path control in software-defined networking (SDN). Each traffic pair is modeled as an agent selecting one of three retained candidate paths, while centralized critics learn coordinated decisions from topology-specific Ryu–Mininet transition datasets. Nine policies are compared using ten paired seeds on fat-tree, mesh-grid, and WAN-corridors topologies under a deployed utilization–latency weighting of 0.60/0.40, together with flow-completion, latency, congestion, architectural-comparison, sensitivity, robustness, statistical, and controller-overhead analyses. The utilization-aware path heuristic achieves the strongest overall reward ranking. MADDPG is the strongest learned policy on fat-tree, is not significantly outperformed by any evaluated policy on mesh-grid, and remains statistically tied with completion-matched policies on WAN-corridors. Behavior adjustment is topology-dependent rather than uniformly beneficial. The exported policy requires approximately 52μs per joint decision, whereas complete control-loop timing is dominated by network-statistics polling. These results support offline multi-agent SDN control as a competitive, low-overhead option when interpreted jointly with topology structure, flow completion, and strong heuristic baselines.
A. Kyzyrkanov, Y. Nurakhov, Zhenis Otarbay et al.· Technologies· 0 citations