Traditional wireless networks are built on a foundation of static mathematical models and rigid, rule-based protocols. Engineers design these systems by optimizing individual layers—such as the physical layer (PHY) for signal transmission or the medium access control (layer) for resource scheduling—using deterministic formulas. However, as next-generation networks expand to support billions of heterogeneous IoT devices, ultra-low latencies, and complex topologies like space-air-ground-sea integration, these classical analytical models become mathematically intractable.
To overcome this complexity, early integrations of machine learning (ML) treated artificial intelligence as an external "overlay"—an auxiliary optimization tool running on top of an already established network fabric. As noted by Li et al. in Series Editorial The Sixth Issue of the Series on Machine Learning in Communications and Networks, this fragmented approach isolates ML within specific layers or tasks, preventing the network from achieving holistic, real-time adaptability. Khan et al. argue in 6G Wireless Systems: A Vision, Architectural Elements, and Future Directions that separating AI from the core architecture introduces latency and administrative overhead, rendering it incapable of managing microsecond-level physical layer dynamics.
The solution is a paradigm shift toward AI-native network architecture. In this framework, machine learning is not an add-on service but a fundamental design requirement embedded directly into the network fabric. As Saad explains in the Inaugural Editorial for TMLCN, modern network complexity requires a ground-up transition from manual, rule-based scripting to autonomous, data-driven management.
The Intelligence Plane and the Closed-Loop System
To understand why an AI-native architecture works, one must look at how it restructures the classic networking model. Traditional networks separate operations into a data plane (which forwards traffic) and a control plane (which decides where traffic goes). An AI-native architecture introduces a third pillar: the intelligence plane.
As detailed by Atov et al. in Data Science and Artificial Intelligence for Communications, the intelligence plane integrates dedicated hardware modules—such as local accelerators and processing units—directly into network switches and base stations. This structural change establishes a localized sensing-cognizing-acting loop inside the network hardware itself.
[ Wireless Environment ]
│ ▲
▼ (Sensing) │ (Acting)
┌──────────────────────────────┐
│ INTELLIGENCE PLANE │
│ (Cognition & Deep Learning) │
└──────────────────────────────┘
Instead of sending telemetry data up to a distant cloud server and waiting for an optimization command, the network node observes its local environment, processes the data through embedded models, and reconfigures its parameters in real-time. Qin et al. demonstrate in 20 Years of Evolution From Cognitive to Intelligent Communications that this perception-action cycle transforms networks from reactive systems into proactive entities. By treating intelligence as a native node within the infrastructure rather than an external application, as argued by Zhou et al. in A Review of Gaps between Web 4.0 and Web 3.0 Intelligent Network Infrastructure, the system minimizes latency and eliminates the computational bottlenecks of traditional centralized management.
Endogenous Intelligence at the Low-Level Infrastructure
The core mechanism that enables this real-time adaptation is endogenous intelligence—the practice of baking machine learning directly into the physical (PHY) and medium access control (MAC) layers.
In a standard network, tasks like channel estimation, beamforming, and interference mitigation rely on fixed formulas that assume idealized environmental conditions. When a signal encounters real-world obstacles or rapid mobility, these assumptions fail. You et al. show in Towards 6G wireless communication networks: vision, enabling technologies, and new paradigm shifts that embedding ML models directly into the PHY and MAC layers allows the network to adapt stochastically to specific local radio environments without relying on static mathematical models.
According to Hong et al. in 6G R&D vision: Requirements and candidate technologies, this low-level integration automates highly complex resource scheduling and beamforming tasks directly within the signal processing fabric. For example, if an unexpected physical blockage interrupts a high-frequency beam, an AI-native physical layer can stochastically adjust its transmission protocols instantaneously to maintain connectivity, rather than waiting for a higher-layer timeout and renegotiation (Alexiou et al.). This shifts machine learning away from high-level user applications down into low-level infrastructure optimization, such as LTE scheduling and Network-on-Chip communication (Guo et al.).
Autonomous Orchestration and Self-Healing
When intelligence is distributed across every layer and node, the network begins to function as a unified, self-sustaining nervous system. This enables autonomous orchestration, where the network manages its own lifecycle, resource allocation, and energy consumption without human intervention.
This capability operates across multiple scales:
- Dynamic Resource Allocation: Instead of relying on manual configurations, the network uses learning-based frameworks to handle fluctuating demands. Zhang notes in A Message and Mission Statement From the New Editor-in-Chief that this data-driven approach provides a scalable solution for managing billions of IoT devices with highly heterogeneous demands.
- Automated Network Slicing: Managing multi-provider network slices manually introduces immense administrative complexity. Zhao et al. demonstrate in An study on the application of blockchain based 5G Technology in the Power IoT that automating resource allocation through self-executing, intelligent scripts allows for the real-time optimization of bandwidth and latency based on current environmental conditions.
- Self-Healing and Energy Optimization: In highly dynamic environments, nodes must balance performance with energy constraints. Saad et al. explain in A Vision of 6G Wireless Systems: Applications, Trends, Technologies, and Open Research Problems that autonomous network management allows nodes to manage their own energy harvesting and topology adjustments.
If a node fails or interference spikes, the surrounding architecture uses cognitive radio principles to autonomously mitigate the interference and reroute traffic (Xiang; Qadir et al.). This transition from human-scripted maintenance to AI-driven autonomic systems allows the network to solve multi-dimensional optimization problems that are far too complex for manual engineering (Zhu et al.; Yang et al.).
By embedding machine learning as a native architectural feature, next-generation networks move away from rigid, pre-programmed behaviors. They become living, distributed computing platforms capable of sensing their environment, optimizing their own infrastructure, and healing their own faults in real-time.