Intent-Based Networking (IBN)

• Intent-Based Networking (IBN) is a network paradigm that translates high-level intents into automated, device-specific policies, enabling scalable and adaptive network control.
• IBN employs a multi-stage lifecycle including intent creation, normalization through NLP and AI/ML, and closed-loop feedback to ensure configuration accuracy and optimal performance.
• Key techniques such as collaborative filtering and formal optimization allow IBN to effectively manage heterogeneous infrastructures and maximize resource utilization.

Intent-Based Networking (IBN) is a network management paradigm that translates high-level, declarative intents into concrete, automated policies and actions across heterogeneous ICT infrastructures. IBN abstracts the complexity of device-level configurations, allowing operators to specify desired outcomes (“what”) without detailing the implementation (“how”). Contemporary research establishes IBN as essential for enabling autonomous, scalable, and adaptive network control, particularly as networks grow in scale, complexity, and heterogeneity.

1. Fundamental Principles and Architectural Components

IBN defines an “intent” as a high-level network goal, often expressed in controlled natural language, e.g., "ensure <200 ms latency for web browsing over the next 24 hours” (Bensalem et al., 2021). The architecture typically includes:

• North-Bound Interface (NBI): Presents CLI, GUI, REST, or voice-based entry points for intent specification, purposely hiding device-level complexity.
• Intent Manager / Parser: Employs regular expressions, CNL grammar, and NLP/LSTM methods to map user intents to internal representations, supported by a knowledge base.
• Policy Configurator / Builder: Matches parsed intents with policy templates and resolves conflicts among overlapping intents.
• Intent Compiler / Translator: Generates device- or controller-specific rules (e.g., OpenFlow, P4, NETCONF/YANG) from high-level abstractions.
• South-Bound Interface (SBI): Enforces generated policies on SDN controllers and network devices (Bensalem et al., 2021).
• Monitoring & Telemetry: Provides closed-loop feedback for compliance checking and automated correction.
• AI/ML Engine: Assists intent extraction, anomaly detection, resource prediction, and adaptive learning.

The pipeline supports dynamic adaptation, auditability, and explainability, aligning with the requirements of scalable, multi-domain environments (Bensalem et al., 2021).

2. Formalization of Intents and Translation Workflows

Intent is typically formalized as a tuple or structured object:

• General Model: $I = (ID, S, O, C, K, ctx)$ where S is stakeholders, O is objectives, C is constraints, K is KPIs/SLOs, and ctx is contextual metadata (Mehmood et al., 2021).
• ICT Supply Chains: $I = \{\,i_k=(U_k,A_k,P_k,C_k)\,\}$, reflecting user, asset, permission, and constraint sets (Bensalem et al., 2021).
• Vehicular Edge Computing: Intents specify joint compute and network requirements, with node and link constraints mapped into substrate resource-embedding problems (He et al., 2023).

Translation involves intent ingestion, normalization (tokenization and mapping), parsing via grammars, and decomposition into actionable intermediate representations (policy graphs, device rules, or configuration descriptors). Formal translation functions, such as $T:\mathcal{I}\rightarrow\mathcal{P}$, map intents to low-level policies subject to conflict-freeness and optimization objectives (Bensalem et al., 2021).

A typical multi-stage lifecycle comprises: creation, normalization, validation, decomposition, rendering/deployment, assurance, and termination (Mehmood et al., 2021).

3. Handling Heterogeneity, Scalability, and Dynamic Adaptation

IBN frameworks are explicitly designed to manage heterogeneous ICT systems, trading off between centralized and decentralized orchestration:

• Heterogeneous Platforms: Devices span CPUs, GPUs, FPGAs, TPUs; ML-based IBN must select and schedule tasks across diverse hardware (Bensalem et al., 2021).
• Collaborative Filtering for Benchmarking: When performance data for new (ML, device) pairs is sparse, collaborative filtering (SVD+SGD factorization) is used to predict inference time or throughput, allowing near-optimal placement and scheduling (~3%–20% normalized RMSE with 30%–90% missing values) (Bensalem et al., 2021).
• Periodic Retraining and Warm-Up: Small budgets of random benchmarks for new devices or models, with periodic retraining, are recommended to maintain accuracy. Explicit side-information (model complexity, device FLOPs) accelerates convergence.
• Resource Mapping in VEC: Formal optimization and heuristic algorithms embed intents as microservice graphs with both compute and network constraints, leading to high utilization (up to 76%), high acceptance ratios (up to 71%), and >95% reduction in orchestration time relative to standard approaches (He et al., 2023).

Closed-loop feedback with real-time telemetry ensures resilient adaptation to failures, mobility, or resource churn.

4. Integration of AI/ML for Automation and Performance Assurance

Machine learning is core to state-of-the-art IBN:

• Intent Parsing: Sequence-to-sequence models, LSTMs, or transformer-based encoders extract and classify intents, supporting NLP-based, multimodal, or voice-driven command channels.
• Policy Placement and Resource Prediction: Latent factor models map application intents to physical resources; inference performance and system cost are estimated for optimal placement (Bensalem et al., 2021).
• Autonomous Benchmarking: Lightweight collaborative filtering approaches allow inference performance to be predicted on unseen (model, device) combinations with sub-10% error, enabling scheduling in environments where exhaustive benchmarking is infeasible (Bensalem et al., 2021).
• Learning for Conflict Resolution: AI/ML modules accelerate intent reconciliation in complex, dynamic multi-intent scenarios, such as overlapping access controls in supply chains (Bensalem et al., 2021).

AI/ML modules close the loop between network state, intent translation, and enforcement, ensuring operational consistency and optimality.

5. Performance Metrics, Evaluation, and Empirical Outcomes

IBN systems and benchmarking frameworks are evaluated using:

Metric	Definition/Formula	Context
Inference Performance Target	Time per sample (e.g., ms/image) in inference mode	ML function selection and placement (Bensalem et al., 2021)
Loss Function	Regularized squared error over benchmarks: $$L(Q,P) = \sum_{(i,j)}(e_{i,j} - q_i^\top p_j)^2 + \lambda (||q_i||^2+||p_j||^2)$$	Performance prediction for unknown (model, device) pairs (Bensalem et al., 2021)
Normalized RMSE	$$\mathrm{nRMSE} = \frac{\mathrm{RMSE}}{(\max R - \min R)}$$, RMSE over held-out predictions	Error metric for performance estimate generalization
Acceptance Ratio ($\mathrm{AR}$)	Long-term average: $$\lim_{T\to\infty} \frac{\sum_{t=0}^T\sum_{m}X^{m,t}}{\sum_{t=0}^T|\{I^{m,t}\}|}$$	Intent fulfillment in VEC (He et al., 2023)
Resource Utilization ($\mathrm{RC}$)	Ratio of fulfilled intent revenue to embedding cost: $$\mathrm{RC} = \lim_{T\to\infty}\frac{\sum_{t=0}^T\sum_m\varepsilon^{m,t}}{\sum_{t=0}^T\sum_m\kappa^{m,t}}$$	Efficiency in network-edge computing orchestration

Empirical benchmarks confirm that collaborative filtering-based performance estimation for ML function deployment yields normalized RMSE below 0.07 in most regimes, and dynamic orchestration in VEC environments achieves resource utilization and intent acceptance far surpassing standard heuristics (Bensalem et al., 2021, He et al., 2023).

6. Deployment Recommendations and Operational Practices

Research identifies the following deployment strategies:

• Warm-Up and Retraining: Whenever a new model or device is brought into the system, allocate a small warm-up budget for benchmarking, and periodically retrain the collaborative filtering model to maintain prediction accuracy as environments evolve (Bensalem et al., 2021).
• Incremental Augmentation: Incorporate explicit side-information—such as model complexity and device compute profiles—to further speed up the cold-start problem.
• Policy Engine Integration: Use predicted performance metrics as inputs to intent scheduling, right-sizing, and resource procurement.
• Closed-Loop Monitoring: Integrate the prediction and learning modules with runtime telemetry for ongoing compliance assurance.
• Generalization: The same collaborative filtering and learning-based scheduling can generalize to deploying non-ML functions (e.g., VNFs, stream analytics) on heterogeneous platforms.

Practitioners are advised to complement standard IBN stacks with lightweight, data-driven benchmarking and AI-driven intent translation layers to realize robust, scalable, and efficient automated network management (Bensalem et al., 2021).