AI/ML-Native MAC Scheduling — Telecom Resource

MAC Overview Scheduling BSR Link Adapt Beamforming CA ISAC NTN MAC Energy-Saving MAC IAB MAC Cell-Free MAC

1. PF vs AI Scheduler

Classical Proportional-Fair balances throughput vs fairness with a fixed heuristic. AI schedulers learn the optimal policy from data — adapting to traffic patterns, channel conditions, and QoS simultaneously.

PF vs AI — Throughput & Fairness Comparison

Compare resource allocation. PF wastes PRBs on suboptimal assignments; AI optimizes globally.

UEs: 8 Scheduler:

Proportional-Fair

Balanced but static. Fixed alpha. Cannot adapt to non-stationary traffic bursts.

Round-Robin

Perfect fairness but ignores channel quality. 30-50% throughput loss.

AI/ML Scheduler

Learns from data. Adapts to traffic + channel. 15-30% throughput gain over PF.

2. Predictive Resource Allocation

AI predicts traffic before BSR arrives and pre-allocates resources. Reduces scheduling latency from 4-10 ms to <1 ms for predictable traffic patterns.

Predictive Allocation — Pre-Allocate Before BSR

Watch AI predict the next packet and pre-allocate. BSR arrives to find resources already ready.

Traffic type:

Latency reduction: Classical: SR (1ms) + BSR processing (2ms) + grant (1ms) = ~4-10ms. Predictive: grant ready before BSR = ~0ms scheduling latency.

3. Reinforcement Learning Training

The AI scheduler learns via RL: observes state, takes action, receives reward (throughput + fairness - latency violations), and updates its policy. Training converges over thousands of episodes.

RL Training Loop — State, Action, Reward, Update

Watch the RL loop. Each episode improves the policy. See cumulative reward converge.

Learning rate:

RL Component	Value in Scheduling
State	CQI, BSR, HARQ, 5QI, PHR per UE
Action	PRB allocation, MCS, beam per UE
Reward	alpha x throughput + beta x fairness - gamma x latency
Algorithm	PPO / SAC (stable, continuous + discrete)

4. Deployment & Fallback

Aspect	Requirement	Solution
Inference latency	<1 ms (slot)	NPU/GPU, model quantization
Model distribution	OTA update	RRC signaling or O-RAN RIC
Fallback	Model failure	Auto-revert to PF (confidence check)
Training data	Diverse scenarios	Simulated + live (hybrid)
Explainability	Regulatory	SHAP/LIME for decisions

Federated learning: Each gNB trains locally on its data, sends only model weights to an aggregator. No raw data leaves the gNB. Global model improves with network scale.

5. Standardization & Field Tips

3GPP Rel-18

AI/ML for NR air interface study item. Defines use cases: beam management, CQI prediction, positioning.

O-RAN RIC

RAN Intelligent Controller with xApps (near-RT) and rApps (non-RT). Deploy ML schedulers as xApps.

NWDAF

Network Data Analytics Function provides analytics to MAC for intelligent scheduling decisions.

6. 6G Vision — AI-Native Air Interface

Learned PHY+MAC

Single neural network: bits in, waveform out. Autoencoder approach. No separate PHY/MAC design.

SON 2.0

Self-optimizing network. RL configures everything. No manual parameter tuning. Continuous learning.

Semantic Comms

MAC understands content meaning. Prioritize important info. 10-100x spectral efficiency for specific tasks.

Scheduling Evolution — Heuristic to AI-Native

Compare scheduling paradigms across generations.

Generation:

Dimension	4G	5G NR	5G-Adv	6G
Scheduler	PF/RR	PF + QoS	ML-assisted	AI-native
Prediction	None	CQI pred.	Traffic pred.	Full predictive
Beam mgmt	N/A	Sweep	AI-predicted	Learned
Link adapt	Table	Table + OLA	ML-MCS	End-to-end
Training	None	Offline	Hybrid	Online + FL