Neural Networks in Network Routing: How AI Optimizes 5G and Beyond

A 5G base station makes hundreds of thousands of resource allocation decisions per second: which frequency band to assign each device, how much bandwidth to grant each connection, which transmission power to use, and how to route traffic through a network that might span dozens of nodes across fiber, millimeter-wave radio, and legacy 4G infrastructure. Traditional algorithms—shortest path routing, proportional fair scheduling, load-balancing heuristics—were designed for more static, predictable networks and strain under 5G's combination of massive device density, heterogeneous traffic (video streaming, autonomous vehicles, IoT sensors, AR/VR), and sub-millisecond latency requirements. Ericsson estimated in a 2024 report that AI-native network management could reduce 5G operating expenses by $100 billion annually across the global telecommunications industry by eliminating inefficiencies in spectrum utilization and reducing manual network tuning.

Key takeaways

• Deep reinforcement learning agents can continuously optimize 5G radio resource allocation, improving spectral efficiency by 15–30% over static algorithms in real deployments.
• Graph neural networks model network topology to enable routing decisions that account for network-wide state rather than local link conditions.
• Network slicing—dividing a physical 5G network into virtualized slices for different applications—requires ML-based admission control to allocate resources dynamically while guaranteeing slice SLAs.
• O-RAN (Open Radio Access Network) architecture is the industry framework enabling third-party AI applications ("rApps" and "xApps") to run ML-based optimization on live network data.

The 5G Resource Allocation Problem

5G networks operate across multiple frequency bands simultaneously—sub-6 GHz for coverage, millimeter wave (24–100 GHz) for capacity, and mid-band (3.5 GHz) for a combination of both. Each base station (gNodeB in 5G terminology) manages spectrum assignments for potentially thousands of simultaneous connections, each with different quality-of-service requirements: a surgical robot requires sub-millisecond latency with zero packet loss; a smart meter needs only kilobits per second with best-effort delivery.

The scheduler inside the base station makes resource allocation decisions every 1 millisecond (one "transmission time interval"). Classical proportional fair scheduling gives each user a share of resources proportional to their channel quality relative to their recent allocation—a greedy algorithm that's simple and efficient but doesn't account for application requirements, future channel predictions, or network-wide load.

The opportunity for machine learning: modern base stations can observe rich contextual information—channel quality indicators from hundreds of devices, historical traffic patterns, neighboring cell loads, application type inferred from traffic classification—that traditional algorithms cannot exploit. An ML-based scheduler trained on this data can learn allocation policies that maximize a complex objective function (throughput, fairness, latency, energy efficiency) in a way that hand-designed algorithms cannot.

Deep Reinforcement Learning for Radio Resource Management

Nokia Bell Labs published results in 2023 from a deep RL system for 5G radio resource management deployed in a live network trial with Deutsche Telekom. The RL agent observes cell-level state (current load, channel conditions, neighboring cell interference) and decides how to adjust transmission power, handover thresholds, and antenna beam configurations. Rewards are defined by a combination of user throughput, latency, and coverage—a multi-objective function that Nokia engineers tuned over several iterations.

In the live trial across 50 base stations in Frankfurt, the RL system improved average user throughput by 22% and reduced dropped call rate by 31% compared to the vendor's default parameter settings, without any increase in transmit power. The key advantage was contextual adaptation: at peak hours, the agent prioritized throughput by tightening handover thresholds (keeping users connected to their current cell longer); during off-peak hours, it widened thresholds to reduce signaling overhead. A human network engineer would need to manually change these parameters twice a day; the agent updated them continuously.

Ericsson's AI research division has published similar results for their "intent-based networking" product, where network operators specify high-level goals ("minimize end-to-end latency for video streaming users in zone A") and an RL agent translates these intents into specific parameter adjustments across the network. Their 2024 commercial deployment data across 10 European operators showed average RAN efficiency improvements of 18%.

Graph Neural Networks for Network-Wide Routing

Beyond the base station scheduler, network routing—deciding which path packets take through the core network—is another domain where classical algorithms (OSPF, BGP, MPLS traffic engineering) leave performance on the table. These algorithms optimize locally based on link state information; they don't account for predicted future traffic, cross-layer interactions, or application-level requirements.

Graph neural networks (GNNs) are particularly well-suited for routing because networks are naturally represented as graphs: nodes are routers or switches, edges are links with capacity and delay attributes, and the routing problem is to find paths through this graph that minimize congestion subject to bandwidth constraints.

Facebook (Meta) published their "DOTE" (Deep Learning for Traffic Engineering) system in 2022, used to optimize traffic distribution across their backbone network. A GNN encodes the current network graph state and predicts how different traffic allocation decisions will affect link utilization across the entire network. A solver then finds the allocation that minimizes maximum link utilization (a proxy for congestion). Meta reported that DOTE reduced peak link utilization by 19% compared to their previous heuristic-based traffic engineering system—significant at Meta's scale, where a 1% improvement in backbone efficiency represents millions of dollars in hardware deferral.

Microsoft Research's 2024 "Reinforcement Learning for Traffic Engineering" work went further, using RL to handle the dynamic version of the problem: routing in real-time as traffic patterns change rapidly. Their system adapted routing configurations at 30-second intervals on a simulated Microsoft Azure backbone, reducing 99th percentile latency by 34% compared to OSPF in scenarios with sudden traffic spikes.

Network Slicing and ML-Based Admission Control

One of 5G's signature capabilities is network slicing: the ability to partition a physical network into multiple isolated virtual networks, each with dedicated resources and guaranteed service levels. An operator might run one slice for URLLC (Ultra-Reliable Low-Latency Communications) serving industrial automation (requires <1ms latency, 99.9999% reliability), another for eMBB (enhanced Mobile Broadband) serving consumer video streaming, and another for mMTC (massive Machine Type Communications) serving IoT sensors.

Admission control—deciding whether to accept a new slice request given current network state—is a combinatorial optimization problem. Accept too aggressively and you violate existing slice SLAs; accept too conservatively and you waste network capacity. ML-based admission control learns to predict the impact of accepting a new slice request on existing slices, enabling more aggressive resource utilization while maintaining guarantees.

Researchers at KTH Royal Institute of Technology published a 2023 paper showing that a deep Q-network (DQN) for slice admission control achieved 94% resource utilization while maintaining SLA compliance, compared to 79% for conservative threshold-based admission control. The improvement comes from the RL agent learning nuanced patterns: a video streaming slice request during off-peak hours can safely be accepted at lower resources than the same request during peak hours.

O-RAN: The Architecture That Makes AI Radio Possible

The most significant infrastructure change enabling AI in 5G networks is O-RAN (Open Radio Access Network), an industry initiative (O-RAN Alliance, members include AT&T, Deutsche Telekom, NTT, China Mobile) to disaggregate and standardize the radio access network architecture. O-RAN introduces two new software platforms for running AI applications:

The Non-Real-Time RIC (RAN Intelligent Controller) operates on timescales greater than 1 second, running "rApps" that perform AI-based analysis, training, and policy optimization. Examples: anomaly detection in network KPIs, training updated RL models and deploying them to the near-RT RIC.

The Near-Real-Time RIC operates at 10ms–1 second timescales, running "xApps" that consume live network data and make resource allocation decisions. Examples: handover optimization, interference management, traffic steering between 4G and 5G.

AT&T deployed their first O-RAN-based AI xApp (a traffic steering xApp that routes users between 4G and 5G based on real-time load predictions) in 2023, reporting a 15% improvement in data session success rate in pilot markets. The open architecture is key: it allows telecom operators to run third-party AI applications (from vendors like Mavenir, Altiostar, or academic research groups) without replacing their entire RAN infrastructure.

FAQ

Q: Why can't traditional algorithms like Dijkstra's shortest path handle modern network routing? A: Dijkstra finds the shortest path by one metric (usually link cost or latency), but doesn't optimize for multiple simultaneous objectives (latency, bandwidth, reliability, energy), can't predict future traffic, and doesn't account for the impact of one routing decision on the rest of the network. Real network routing requires balancing complex multi-objective functions under uncertainty, which ML systems handle better.

Q: How does an RL agent learn to optimize a live network without causing outages during training? A: Production deployments use a "safe RL" approach: the agent learns from a digital twin (a simulation of the network trained on historical data) before any live deployment. In production, it operates within predefined safety envelopes (no action can reduce network performance below a fallback threshold), and a supervisory system monitors for anomalies and reverts to baseline if triggered.

Q: What data does a 5G base station collect that ML can use for optimization? A: Channel Quality Indicators (CQI) reported by every connected device every millisecond, Reference Signal Received Power (RSRP) and quality (RSRQ), buffer status reports from devices, per-user throughput and latency measurements, inter-cell interference measurements, and traffic classification from deep packet inspection (for traffic classified as video, gaming, or IoT). This creates thousands of input features per cell per second.

Q: Does AI-based network optimization work with encrypted traffic? A: Yes, because most optimization decisions are based on network-layer metadata (packet size distributions, inter-arrival timing, connection duration, source/destination address patterns) that are observable even when payload content is encrypted. Traffic classification ML systems can identify application types (video streaming, voice, gaming) from these metadata features without inspecting payload content.

Q: What's the relationship between AI network optimization and network energy efficiency? A: Significant. ML-based load balancing and sleep mode management—shutting down underutilized cells during low-traffic periods and waking them as demand increases—can reduce 5G RAN energy consumption by 20–30%. Ericsson's Smart Energy Saving feature, which uses ML to predict traffic patterns and activate/deactivate cell components accordingly, has been deployed across hundreds of operators and is now responsible for billions of kWh of annual savings globally.

---

Read next:

• How AI Designs Its Own Chips: Hardware-Aware Neural Architecture Search
• AI Compilers and Code Generation: How ML Is Reinventing Software Optimization
• How Image Recognition Works: The Science of Convolutional Neural Networks