The proposed SDN-Expert architecture ( Figure 1 ) consists of three distinct and complementary layers:

1. Application layer

Interacts directly with users, capturing specific expectations and requirements related to network performance and energy management.

2. Control layer

This layer forms the core of our innovative approach, integrating an artificial intelligence-based expert system. It comprises a hybrid inference engine, explicitly combining decision trees and neural networks, along with a rule base and a fact base to ensure optimized adaptive decision-making in real time.

3. Infrastructure layer

Includes physical network equipment (routers, switches) executing the adaptive decisions made by the Control layer, such as dynamic activation and deactivation of router ports based on real-time network conditions.

The significant improvement introduced by our model is the explicit integration of hybrid artificial intelligence techniques (decision trees and neural networks), which ensures greater adaptability and responsiveness compared to traditional, non-adaptive methods.

The problem is formulated as an integer linear optimization problem (ILP). Our formulation improves upon standard ILP models by explicitly incorporating adaptive decision-making mechanisms through AI-based inference, particularly using decision trees for quick decision paths and neural networks for predictive capabilities, thus enhancing responsiveness to rapid changes in traffic conditions.

The objective is to minimize total energy consumption by dynamically switching off unused router ports, while strictly adhering to Quality of Service (QoS) constraints:

Objective function:

$Min\underset{i}{{\displaystyle \sum }}{C}_i\left(t\right)-g_i\left(t\right)$ (1)

Subject to the following constraints:

$L_i\left(t\right)\le L_{max}$ (2)

$\underset{iϵN}{{\displaystyle \sum }}{P}_i\left(t\right)\ge P_{min}$ (3)

$\underset{iϵF}{{\displaystyle \sum }}{B}_i\left(t\right)\ge B_{req}$ (4)

$E_{tot}\left(t\right)\le E_{budget}$ (5)

Where:

$C_i\left(t\right)$: total consumption of the router $i$ at time $t$.

$g_i\left(t\right)$: Save energy by disabling unused ports [7] .

$L_i\left(t\right)$: router latency $i$ at time $t$.

$L_{max}$: maximum acceptable latency threshold.

$P_i\left(t\right)$: number of active ports on the router at the moment $t$.

$P_{min}$: minimum number of active ports to guarantee connectivity.

$B_i\left(t\right)$: available bandwidth for critical flows.

$B_{req}$: minimum bandwidth required to guarantee quality of service for critical flows.

$E_{tot}\left(t\right)$: total energy consumption of the network at time $t$.

$E_{budget}$: maximum energy consumption limit.

These constraints ensure that the solution optimizes energy consumption while maintaining network performance at acceptable levels.

These constraints collectively ensure that the SDN-Expert solution effectively balances energy optimization with maintaining optimal network performance, thus clearly improving upon existing approaches.

The following algorithm follows directly from the proposed model:

Algorithm 1. SDN-Expert algorithm.

To comprehensively assess the performance of the proposed SDN-Expert architecture, we conducted extensive simulations on a realistic network topology consisting of 50 nodes, utilizing real-world traffic traces from the CAIDA dataset. The results obtained were compared to existing methods: Paramasivam & Velusamy (2023), Qi et al. (2019), and Hussein et al. (2024). Performance was analyzed based on three critical criteria: energy savings, average latency, and packet loss rate.

The simulation results clearly illustrate that the SDN-Expert approach achieves a significant energy saving of 32%, compared to 18% for Paramasivam (2023), 12% for Qi (2019), and 20% for Hussein (2024), as shown in Figure 2 .

This substantial improvement primarily stems from integrating a hybrid inference engine that combines decision trees and neural networks, enabling optimized real-time adaptive management of network resources.

Regarding latency, our results show that SDN-Expert achieves a notably lower average latency (40 ms), clearly outperforming existing methods (55 ms for Paramasivam, 65 ms for Qi, and 50 ms for Hussein), as illustrated in Figure 3 .

This improvement is attributed to the AI-based inference engine, which dynamically optimizes network paths by proactively predicting traffic variations and avoiding congestion areas.

The related Figure 4 demonstrates that our model significantly reduces the packet loss rate to only 1%, compared to 3% for Paramasivam, 5% for Qi, and 2.5% for Hussein.

This notable improvement results from predictive mechanisms provided by the neural network and dynamic decision-making facilitated by decision trees, allowing the network to proactively prevent congestion in real time.

This comprehensive comparative analysis clearly demonstrates that the SDN-Expert approach significantly improves SDN network performance regarding energy optimization and Quality of Service (QoS), ensuring high adaptability to real-time variations in network traffic. The explicit integration of decision trees and neural networks grants our model enhanced responsiveness to rapid changes in actual network conditions. As shown in Table 1 , the SDN-Expert approach significantly improves SDN network performance in terms of energy optimization and QoS.