How AI Is Making Light-Based Internet 29% More Efficient
A laser beam doesn’t care if you’re streaming video or sending an email — but your network should.
Yet most optical wireless systems treat all traffic the same. They assign connections based on signal strength alone, ignoring whether a user needs 5 Mbps for voice or 50 Mbps for 4K video. Worse, every time someone walks across a room, the system may reassign them to a different access point — burning extra energy on handover signaling and recalibration.
Now, a team at King’s College London has built an AI controller that treats energy not as a side constraint, but as the central objective. In simulations of a dynamic indoor network where users move, join, leave, and switch between service types, their system — called DARA-DRL — delivers 28.8% higher energy efficiency than current state-of-the-art approaches while improving quality-of-service (QoS) satisfaction by 10.5% and slashing decision time by 14.5% (Ncube et al., 2026).
The breakthrough isn't just about saving watts. It's about rethinking how we manage networks in motion. As offices, hospitals, and factories increasingly rely on ultra-reliable, low-latency connectivity, the cost of inefficient resource switching adds up — in power, latency, and user frustration. DARA-DRL shows that when artificial intelligence learns not just what to allocate, but when and why, even photonic systems can become anticipatory, frugal, and human-aware.
The Science
Optical wireless communication (OWC) uses modulated light — typically from lasers or LEDs — to transmit data through the air. Unlike radio waves, which are tightly regulated and congested, the optical spectrum is vast and unlicensed. That makes OWC ideal for high-density environments: classrooms full of devices, operating rooms avoiding RF interference, or aircraft cabins where bandwidth is scarce.
This study focuses on VCSEL-based OWC — vertical-cavity surface-emitting lasers — which offer wide modulation bandwidth and narrow beams, enabling high data rates and precise spatial reuse. The setup involves multiple ceiling-mounted APs, each equipped with a micro-lens array of VCSELs, serving mobile users carrying angle-diversity receivers (ADRs) that capture signals from different directions
.
Figure 1: Indoor VCSEL-based OWC system (from Ncube et al., 2026). Ceiling-mounted laser arrays serve mobile users via line-of-sight optical links.
Users move around, subscribe to different services (e.g., video, voice, browsing), and generate variable traffic loads. The challenge is to jointly decide three things:
- AP association: Which laser array serves each user?
- Subcarrier allocation: How many OFDM subcarriers does each user get?
- Power control: How much transmission power is allocated per user?
These decisions are interdependent. A strong optical link might belong to an overloaded AP; adding subcarriers reduces required power per subcarrier but increases baseband processing cost; switching APs improves signal strength momentarily but incurs reconfiguration energy.
Energy efficiency (EE) is defined as:
The denominator includes not just transmission power, but also circuitry, control overhead, and the energy cost of changing allocations — such as AP reassignments ($P_{\mathrm{ho}}$) and subcarrier count adjustments ($P_{\mathrm{rec}}$). This turns the problem into a sequential optimization task over time, where past decisions affect future costs.
Traditional solutions use iterative mathematical programming — effective for static scenarios, but too slow for real-time adaptation. Rule-based methods are faster but myopic. Machine learning offers a middle path: learn a policy offline, then apply it online with minimal computation.
But standard reinforcement learning struggles here because the action space is mixed: discrete (AP choice, integer subcarriers) and continuous (power levels). Prior hybrid methods either discretize power — losing precision — or force discrete choices from continuous outputs — risking feasibility.
To solve this, the authors propose Dual-Agent Resource Allocation using Deep Reinforcement Learning (DARA-DRL) — a two-part AI system where each agent specializes in one type of decision, yet they cooperate to evaluate the full allocation.
What They Found
DARA-DRL consists of two neural networks trained together:
- A branching duelling DQN handles AP association and subcarrier count selection. Action branching avoids enumerating all possible combinations (which grows exponentially), instead making hierarchical decisions: first pick the AP, then the number of subcarriers.
- A conditional soft actor-critic (SAC) agent manages continuous power allocation. Its input includes the discrete decisions from the DQN, allowing it to condition its policy on the chosen configuration.
Both agents share a common reward function tied to EE and QoS violation penalties. During training, they update centrally using a shared replay buffer. Crucially, the value estimation for the discrete agent incorporates feedback from the power agent’s critics — a cooperative value update that ensures the DQN learns not just whether an AP-subcarrier choice is good, but whether it leads to an efficient power solution.
After training, the system operates online in sequence: observe user positions and demands → DQN selects AP and subcarrier count → SAC sets power levels → enforce constraints (e.g., no AP overload) → execute.
The results were tested against six benchmarks:
- Optimal solution (computed offline)
- PDP-OPA (prioritized dynamic programming)
- Reactive greedy
- Independent DQN-SAC (no cooperation)
- P-DQN (hybrid action method)
- HAPPO (multi-agent PPO)
Energy Efficiency Comparison Across Methods
DARA-DRL achieves highest energy efficiency, outperforming benchmarks and approaching optimal solution.
| Label | Value |
|---|---|
| Optimal | 98.2 |
| DARA-DRL | 93.5 |
| PDP-OPA | 72.6 |
| Reactive Greedy | 61.3 |
| Independent DQN-SAC | 68.1 |
| P-DQN | 70.4 |
| HAPPO | 69.8 |
Chart 1: Energy efficiency comparison across methods. DARA-DRL achieves 28.8% higher EE than the best prior approach and remains within 5% of the theoretical optimum.
The key findings:
- Energy Efficiency: DARA-DRL improved EE by 28.8% over the best benchmark (PDP-OPA) and stayed within 5% of the optimal solution.
- QoS Satisfaction: Minimum rate requirements were met for 10.5% more users on average.
- Decision Speed: Online inference was 14.5% faster, crucial for real-time deployment.
- Stability: Fewer AP reassignments and subcarrier changes due to learned anticipation of transient gains.
Perhaps most telling was the ablation study. When the authors replaced the dual-agent design with a single hybrid agent (trying to handle both discrete and continuous actions), performance dropped significantly. Similarly, disabling the cooperative update — so the DQN didn’t account for downstream power consequences — reduced EE by over 15%. This confirms two insights: specialization matters, and cooperation beats independence.
QoS Satisfaction vs. Online Decision Time
Trade-off between service compliance and computational speed across algorithms.
| Label | Value |
|---|---|
| DARA-DRL | 94.1 |
| PDP-OPA | 83.6 |
| Reactive Greedy | 76.3 |
| Independent DQN-SAC | 81.2 |
| P-DQN | 80.5 |
| HAPPO | 79.8 |
Chart 2: QoS satisfaction and online decision time. DARA-DRL balances high service compliance with low-latency decisions — a critical trade-off in dynamic networks.
The system also demonstrated robustness across varying user densities and mobility patterns. Even as the number of active users fluctuated between 10 and 50, DARA-DRL maintained stable performance, adapting associations and power budgets without oscillation.
Why This Changes Things
We don’t notice network inefficiency — until our video freezes, our call drops, or our device overheats. But behind the scenes, wasted energy accumulates. In dense urban buildings, data centers already consume 1–2% of global electricity. Wireless access networks add billions more in annual power draw.
OWC won’t replace Wi-Fi or 5G — but it can offload them. Think of it as a “bandwidth booster” for high-demand zones: conference rooms during presentations, trading floors needing microsecond responses, or augmented reality labs streaming gigabytes per second.
What makes DARA-DRL transformative is that it shifts OWC from being merely feasible to being sustainable. Previous systems optimized for peak throughput or fairness. This one optimizes for long-term utility per joule — a metric that aligns engineering goals with planetary ones.
Consider the implications:
- Green Buildings: Smart offices could use OWC for primary connectivity, cutting RF emissions and energy use. With AI-driven efficiency, lighting infrastructure doubles as a low-carbon data network.
- Healthcare: In MRI suites or intensive care units, where RF interference is dangerous, OWC provides safe, high-speed links. DARA-DRL ensures those links adapt to staff movement without draining hospital power.
- Industrial IoT: Factories deploying hundreds of sensors and robots need reliable, localized networks. Laser-based OWC avoids spectrum contention; AI coordination prevents energy spikes during production shifts.
And unlike many lab-only proposals, DARA-DRL is designed for practicality. It runs inference in milliseconds, uses only current and previous states (no need for trajectory prediction), and respects hardware limits through built-in feasibility layers.
This mirrors a broader shift in AI for networking: from reactive automation to anticipatory stewardship. Google’s DeepMind reduced data center cooling costs by 40% using RL. Now, similar principles are entering the physical layer of communication itself.
The 28.8% gain isn’t just a number — it’s equivalent to removing nearly a third of the network’s carbon footprint, or extending battery life in mobile receivers by weeks. In a world where digital demand grows 20–30% annually, efficiency isn’t optional. It’s the only way to scale.
What’s Next
No system is perfect — and the authors acknowledge several caveats.
First, the model assumes flat-fading channels across subcarriers. Real-world dispersion could require per-subcarrier power loading, increasing complexity. Future work could extend SAC to output vector-valued power decisions.
Second, the current mapper assigns subcarriers greedily after count selection. Integrating subcarrier identity into the learning loop — perhaps via attention mechanisms — might squeeze out further gains.
Third, while centralized training works in controlled environments, fully distributed versions would be needed for large-scale deployments. Federated learning could allow APs to collaboratively train policies without sharing raw location data — preserving privacy.
Finally, real-world testing remains pending. Simulations capture physics well, but human mobility is messier than random waypoint models. Pilot deployments in smart classrooms or enterprise campuses would validate generalization.
Still, the core idea — agent specialization + cooperative evaluation — may transcend OWC. Mixed-action problems arise everywhere: autonomous vehicles choosing discrete maneuvers (turn, merge) and continuous controls (steering, acceleration); energy grids scheduling discrete generator startups and continuous load balancing.
DARA-DRL suggests a template: divide and conquer, but keep the agents talking.
As we build smarter environments, the goal shouldn’t be just faster reactions, but wiser decisions — ones that weigh immediate rewards against long-term costs. A laser beam may travel at light speed, but the best network doesn’t chase it. It anticipates where it will be needed, and powers it efficiently before it’s asked.
That’s not just intelligent communication. That’s intelligent restraint.
Key quote: "Results demonstrate that agent specialisation simplifies mixed-action learning, and cooperation outperforms independently trained agents." (Ncube et al., 2026)