⚡ electrical engineering

EARL: Energy-Aware Adaptive Antenna Control with Reinforcement Learning in O-RAN Cell-Free Massive MIMO Networks

This paper proposes EARL, a reinforcement learning-based framework for O-RAN cell-free massive MIMO networks that dynamically adapts antenna configurations to minimize total energy consumption while meeting performance requirements, achieving up to 81% power savings within near-real-time constraints.

Original authors: Zilin Ge, Ozan Alp Topal, Irshad Ahmad Meer, Pei Xiao, Cicek Cavdar

Published 2026-02-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zilin Ge, Ozan Alp Topal, Irshad Ahmad Meer, Pei Xiao, Cicek Cavdar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive concert hall where hundreds of tiny speakers (Radio Units, or RUs) are scattered everywhere to make sure every single person in the audience (the Users) hears the music perfectly. This is what Cell-Free Massive MIMO is: a super-powerful 5G/6G network where many small antennas work together to give everyone a great signal.

But here's the problem: Keeping all these speakers turned on at full volume, even when the hall is empty or only a few people are there, is a huge waste of electricity. It's like leaving every light in a stadium on just because one person is watching a game.

This paper introduces a smart solution called EARL (Energy-Aware Adaptive Antenna Control with Reinforcement Learning). Think of EARL as a super-intelligent, energy-conscious DJ who sits in the control booth (the Cloud) and decides exactly which speakers need to be on, and how loud they should be, in real-time.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Always-On" Waste

In traditional setups, the system often keeps all antennas active just to be safe. This is like keeping a fleet of delivery trucks idling in a parking lot, waiting for a single package. It burns fuel (electricity) and costs a lot of money, even if no one needs a delivery right now.

2. The Solution: The "Smart DJ" (EARL)

EARL uses a type of Artificial Intelligence called Reinforcement Learning. Imagine a video game character that learns by playing thousands of times.

The Goal: The character wants to win (keep the music playing clearly for everyone) while losing the least amount of "energy points" possible.
The Learning: At first, the AI might turn on too many speakers. But every time it wastes energy, it gets a "penalty." Every time it saves energy while still keeping the music clear, it gets a "reward."
The Result: Over time, the AI learns the perfect dance. It knows exactly which speakers to turn off when the crowd is small, and which ones to boost when the crowd gets loud, all without anyone telling it what to do.

3. The "Centralized Brain"

The paper highlights a specific setup called O-RAN. Think of this as moving the "brain" of the operation from the speakers themselves to a central cloud computer.

Old Way: Each speaker tries to figure out its own volume and direction locally. This is messy and inefficient.
New Way (EARL): The central cloud sees the whole picture. It knows exactly where every person is standing and what the air conditions are. It can coordinate the speakers like a symphony orchestra, ensuring they work together perfectly. This allows for much more precise energy saving.

4. The Magic Trick: "Greedy Refinement"

The AI makes a great first guess in about 0.2 seconds (faster than a blink!). But the researchers added a second step called "Greedy Refinement."

Analogy: Imagine the AI turns on 10 speakers to be safe. The "Greedy Refinement" is like a meticulous editor who says, "Wait, we only need 5 of those. Let's turn off 3 more and see if the music is still clear."
This second step takes a bit longer (about 2 seconds), but it cuts the energy usage in half again.

5. The Results: Saving the Planet (and the Bill)

The paper tested this system and found amazing results:

Compared to leaving everything on: EARL saved up to 81% of the energy. That's like turning off 8 out of 10 lights in a room and still seeing perfectly.
Compared to "dumb" smart systems: It saved 50% more energy than other existing methods.
Speed: It makes these decisions in the blink of an eye, well within the speed limits required for modern 5G/6G networks.

Summary

EARL is a smart, learning system that acts like a bouncer and a DJ for a massive network of antennas. Instead of keeping the whole club open and loud all the time, it dynamically opens just the right doors and turns on just the right lights to keep the party going for the users, while saving a massive amount of electricity. It proves that we can have high-speed, high-quality internet without burning through our energy budget.

1. Problem Statement

The paper addresses the high energy consumption inherent in Cell-Free Massive MIMO (CF-mMIMO) networks, particularly when deployed within the Open Radio Access Network (O-RAN) architecture.

The Challenge: While CF-mMIMO offers uniform high performance via distributed Radio Units (RUs) and centralized cloud processing, it incurs significant energy costs due to joint transmission, dense RU deployment, and fronthaul traffic.
Limitations of Existing Solutions:
- Optimization-based methods: Computationally intensive and impractical for real-time deployment.
- Existing Machine Learning (ML) approaches: Often focus on RU-level shutdown or downlink power control but neglect antenna-level granularity. Furthermore, they often assume distributed precoding (local interference cancellation), which is less spectrally efficient than the centralized precoding enabled by O-RAN split options 7.1 and 8.
The Goal: To minimize end-to-end power consumption (radio, optical fronthaul, and cloud processing) while strictly meeting user Spectral Efficiency (SE) requirements, specifically under a centralized precoding architecture.

2. Methodology: The EARL Framework

The authors propose EARL (Energy-Aware Reinforcement Learning), an adaptive antenna control framework implemented as an xApp within the O-Cloud (near-Real-Time RIC).

A. System Model & Power Consumption

Architecture: A downlink CF-mMIMO system with $L$ O-RUs connected to a central O-Cloud via lossless fiber fronthaul.
Functional Splits: The model considers O-RAN Split 8 (all processing in Cloud) and Split 7.1 (low-PHY in RU, high-PHY in Cloud).
Power Model: The total power ( $P_{tot}$ $P_{t o t}$ ) is a function of:
1. O-RU Power: Static power per active antenna chain, load-dependent transmit power, and RU-side processing power (dependent on the functional split).
2. O-Cloud Power: Baseband processing (precoding, modulation, coding) and cooling.
3. Fronthaul Power: Optical transceiver power (active only if RU is on) and traffic-dependent cloud-side fronthaul power.
Key Insight: Power consumption scales primarily with the number of active antenna elements and active O-RUs.

B. Problem Formulation

The authors formulate an end-to-end power minimization problem subject to SE constraints.

Objective: Minimize a weighted sum of active antennas and active O-RUs (approximating total power).
Constraints: Ensure the effective Signal-to-Interference-plus-Noise Ratio (SINR) for every User Equipment (UE) meets a minimum threshold.
Complexity: The problem is non-convex and combinatorial because the effective SINR cannot be easily expressed as a closed-form function of the discrete antenna activation variables under centralized precoding.

C. Reinforcement Learning Solution

To solve the combinatorial problem in real-time, EARL uses Proximal Policy Optimization (PPO), a state-of-the-art on-policy RL algorithm.

State Space ( $s_t$ ): Includes the large-scale channel gain matrix ( $\Phi$ ), current antenna activation status ( $n_t$ ), normalized total power, and the fraction of users violating SE constraints.
Action Space ( $a_t$ ): An incremental vector $\{-1, 0, +1\}$ for each RU, adjusting the number of active antennas.
Reward Function ( $r_t$ ): Designed to penalize high power consumption and SE violations. It includes an adaptive Lagrangian multiplier to balance energy efficiency against reliability.
Inference Strategy:
1. RL Decision: The trained PPO agent predicts an antenna configuration based on channel gains (without needing instantaneous CSI).
2. Greedy Refinement: A post-processing step iteratively removes redundant active antennas if the SE constraint remains satisfied, further optimizing the solution.

3. Key Contributions

Comprehensive Power Model: Developed an end-to-end power consumption model for O-RAN CF-mMIMO under Split 7.1 and 8, explicitly linking power to active antenna ports and functional splits.
RL-Based Control: Formulated the antenna activation problem as an RL task, overcoming the computational intractability of traditional optimization methods for real-time control.
Hybrid Approach: Introduced a Greedy Refinement stage that combines the speed of RL with the precision of local search, ensuring constraints are met while maximizing energy savings.
Centralized Precoding Focus: Unlike prior works, this framework is designed specifically for centralized precoding (O-Cloud), leveraging its superior spectral efficiency to enable more aggressive antenna deactivation.

4. Numerical Results

Simulations were conducted in a 400m $\times$ 400m area with 16 O-RUs and varying numbers of UEs (4, 6, 8).

Energy Savings:
- Compared to a "Full-On" baseline (all antennas active): EARL achieves up to 81% power reduction.
- Compared to a Heuristic baseline (channel-strength based activation): EARL achieves up to 50% power reduction.
- The Greedy Refinement step alone cuts the power consumption of the RL-only solution by an additional 50%.
Spectral Efficiency: The framework successfully maintains SE requirements. While the RL-only agent tends to be conservative (higher SE, higher power), the greedy refinement brings SE closer to the threshold, optimizing the trade-off.
Runtime Performance:
- RL-Only: ~220 ms (well within the O-RAN near-RT limit of 1–2 seconds).
- RL + Greedy: ~2.0 seconds (meets the 2-second limit, suitable for slightly less dynamic but highly optimized scenarios).
- Heuristic: ~70 ms (fastest, but significantly less energy-efficient).
Split Comparison: Split 8 (fully centralized) was found to be more energy-efficient overall (30% reduction vs. Split 7.1) due to efficient resource sharing in the cloud, despite higher cloud processing loads.

5. Significance

Feasibility of Green 6G: The paper demonstrates that AI-driven, adaptive antenna control is feasible within the strict latency constraints of O-RAN, offering a practical path to sustainable 6G networks.
Granularity: By operating at the antenna-element level rather than just the RU level, EARL unlocks energy savings (up to 40% more than RU-level control) that were previously inaccessible.
Scalability: The use of PPO allows the system to generalize across different network loads and user distributions without re-optimizing from scratch, making it suitable for dynamic 6G environments.
Architectural Insight: The results validate that centralizing processing (Split 8) combined with intelligent antenna control yields the best balance of spectral efficiency and energy consumption.

In conclusion, EARL provides a robust, low-latency solution for minimizing the carbon footprint of next-generation wireless networks without compromising user quality of service.