AI-Empowered Resource Allocation for Wirelessly Powered Pinching-Antenna Systems

This paper proposes a deep reinforcement learning-based algorithm to jointly optimize antenna positioning, transmit power, and time-switching ratios in a multi-user wirelessly powered NOMA system equipped with pinching antennas, thereby significantly enhancing energy efficiency in dynamic environments compared to conventional fixed-antenna schemes.

The Gist

The Big Picture: A Self-Healing, Energy-Hungry Network

Imagine a future where your smart home, your car, and your wearable devices are all connected, but they are running on tiny batteries that are hard to replace. To solve this, we want these devices to "eat" energy from the air (like solar panels, but using radio waves) before they talk to each other.

This paper proposes a new way to manage a network of these devices. It combines three big ideas:

1. Wireless Charging: Devices harvest energy from a central tower before sending data.
2. NOMA (The "Party Line"): Instead of giving each device its own private phone line, everyone talks at the same time on the same frequency, but at different volumes. The central tower is smart enough to untangle the voices.
3. Pinching Antennas (The "Magic Slider"): Instead of having a fixed antenna, the system uses a long, flexible waveguide with a "pinch" that can slide back and forth to find the perfect spot to send or receive signals.

The Problem: The "Static" vs. The "Dynamic"

The Old Way (Fixed Antennas):
Imagine a lighthouse with a light fixed in one spot. If a ship (your device) moves behind a hill or a building, the light can't reach it. The signal gets blocked, the ship can't charge its battery, and it can't send a message. In a busy city with lots of obstacles, this is a disaster.

The New Way (Pinching Antennas):
Now, imagine that lighthouse has a flexible arm with a light that can slide up and down a pole. If a ship is blocked, the light slides to a new angle to peek over the obstacle. This is what Pinching Antennas (PAs) do. They can physically move their "active" point along a waveguide to dodge obstacles and find the clearest path to your device.

The Challenge: The "Blind Chef"

The system has a huge job to do, but it's playing a game with incomplete information:

• The Battery: The system doesn't know exactly how much battery is left in your device (maybe the sensor is slightly off).
• The Location: It doesn't know your exact location (maybe you moved slightly).
• The Trade-off: It has to decide: Should I spend 50% of the time charging your battery and 50% talking? Or 80% charging and 20% talking? And Where should I slide the antenna?

If it guesses wrong, the device runs out of power, or the data gets lost. Doing the math to find the perfect answer is incredibly hard because the variables are all mixed up (non-convex). It's like trying to solve a Rubik's cube while wearing blindfolded gloves.

The Solution: The "AI Coach" (Deep Reinforcement Learning)

Instead of trying to solve the math equations perfectly (which takes too long), the authors built an AI Coach using a technique called Deep Reinforcement Learning (DRL).

Think of the AI as a video game player who is learning to play a complex strategy game:

1. Trial and Error: The AI tries different strategies (e.g., "Slide the antenna left, charge for 30% of the time").
2. The Score: If the devices get enough energy and send data successfully, the AI gets a high score (Reward). If a device runs out of battery or fails to send data, the AI gets a penalty.
3. Learning: Over thousands of tries, the AI learns a "feel" for the game. It doesn't need to know the exact physics of the radio waves; it just learns that "When the battery is low and the signal is weak, sliding the antenna to the right works best."

How It Works in Practice

1. Harvest Phase: The central tower sends out radio waves. The devices "eat" this energy to charge their batteries. The AI slides the antenna to the best spot to make sure everyone gets a full meal.
2. Transmit Phase: The devices talk back using the NOMA method (everyone talking at once). The AI adjusts the volume (power) of each device and the antenna position to make sure the tower can hear everyone clearly, even the quiet ones.
3. Adaptation: If a user moves or the battery reading is fuzzy, the AI adjusts instantly. It's like a jazz musician improvising when the band changes the tempo, rather than reading a rigid sheet of music.

The Results: Why It Matters

The researchers ran simulations to see how this new system compares to the old "fixed antenna" systems.

• Efficiency: The new system is much more energy-efficient. It wastes less power and gets more data done per unit of energy.
• Flexibility: As they added more "sliding" antennas, the system got better at finding clear paths, though there is a point of diminishing returns (too many antennas cost too much power to run).
• Robustness: Even when the AI didn't know the exact battery level or location of the users, it still performed better than the old static systems.

The Bottom Line

This paper suggests that the future of wireless networks shouldn't rely on static, rigid towers. Instead, we should use sliding, adaptable antennas controlled by smart AI. This allows the network to "dance" around obstacles, charge devices efficiently, and keep everyone connected, even in a chaotic, moving world. It's the difference between a statue trying to wave at you and a person who can run over to get a better view.

Technical Summary

1. Problem Statement

The paper addresses the challenge of maximizing Energy Efficiency (EE) in a multi-user uplink Non-Orthogonal Multiple Access (NOMA) network where user devices (UDs) are energy-constrained and rely on Wireless Power Transfer (WPT) from a Base Station (BS).

Key Challenges:

• Dynamic Environments: Traditional fixed-antenna systems struggle with time-varying channel conditions, shadowing, and non-line-of-sight (NLoS) propagation, especially in high-frequency bands (mmWave/THz).
• Resource Coupling: The system requires the joint optimization of three tightly coupled variables:
  1. Antenna Positioning: The physical location of active radiation points.
  2. Transmit Power Control: Power allocation for uplink NOMA.
  3. Time-Switching Ratio: The fraction of time allocated for energy harvesting vs. data transmission.
• Uncertainties: The optimization must account for imperfect knowledge of user battery states and spatial locations.
• Non-Convexity: The problem is non-convex due to nonlinear energy harvesting (EH) characteristics, coupled variables, and stochastic constraints, making analytical solutions intractable.

2. System Model

The proposed system utilizes a Pinching-Antenna (PA) array mounted on a dielectric waveguide.

• Architecture: Instead of fixed antennas, the BS has a waveguide of length $L$ with $N$ PAs. Each PA can be dynamically activated at any position $x_n$ along the waveguide.
• Protocol: A Harvest-Then-Transmit protocol is used:
  • Phase I (WPT): The BS transmits energy for a duration $\beta(t)T_s$. UDs harvest energy using a practical nonlinear EH model (rectifier saturation effects).
  • Phase II (Transmission): UDs transmit data to the BS using power-domain NOMA for the remaining $(1-\beta(t))T_s$. Successive Interference Cancellation (SIC) is used at the BS.
• Channel Model: The channel gain is a coherent superposition of signals from all active PAs, incorporating waveguide propagation losses, coupling coefficients, and free-space path loss.
• Uncertainties: The model explicitly includes bounded estimation errors for battery levels (uniform distribution) and user locations (Gaussian distribution).

3. Methodology

To solve the non-convex joint optimization problem under uncertainty, the authors propose a Deep Reinforcement Learning (DRL) framework, specifically using the Deep Deterministic Policy Gradient (DDPG) algorithm.

• Markov Decision Process (MDP) Formulation:
  • State Space ($s_t$): Includes the estimated battery levels and spatial locations of all $K$ users.
  • Action Space ($a_t$): A continuous vector comprising the transmit power of all users, the activation positions of all $N$ PAs, and the time-switching ratio $\beta(t)$.
  • Reward Function ($r_t$): Defined as the system Energy Efficiency (Total Rate / Total Power). A penalty term is applied if the minimum rate requirement ($R_{min}$) for any user is violated.
• Algorithm Design:
  • Actor-Critic Architecture: The Actor network ($\pi_\phi$) generates deterministic actions based on the state, while the Critic network ($Q_\theta$) evaluates the state-action pairs.
  • Handling Constraints: Physical constraints (e.g., minimum antenna spacing, battery causality) are enforced through action space normalization and environment-level feasibility checks.
  • Training: The agent learns an optimal policy to maximize cumulative expected rewards without requiring an explicit environmental model or perfect channel state information (CSI).

4. Key Contributions

1. Novel System Architecture: Integration of Pinching-Antenna (PA) arrays into WPT-assisted NOMA networks. This allows for dynamic spatial reconfigurability to establish stronger Line-of-Sight (LoS) paths and mitigate blockage without increasing hardware complexity.
2. Joint Optimization Framework: Formulation of a comprehensive problem maximizing EE by jointly optimizing antenna positioning, power control, and time-switching, while accounting for realistic nonlinear EH and hardware constraints.
3. Robustness to Uncertainty: Development of a DRL-based solution that autonomously learns resource allocation policies despite uncertainties in user locations and battery states, eliminating the need for perfect CSI.
4. Algorithmic Innovation: Adaptation of the DDPG algorithm for continuous action spaces in a highly coupled, constrained communication environment, demonstrating superior convergence and stability compared to other DRL baselines (SAC, A2C, PPO).

5. Simulation Results

Extensive simulations were conducted at 28 GHz with $N=3$ PAs and $K=3$ users.

• Convergence: The proposed DDPG algorithm demonstrated faster convergence and higher stability (lower variance) compared to other DRL algorithms (SAC, A2C, PPO), achieving the highest average reward.
• Energy Efficiency Gains:
  • vs. Fixed Antennas: The PA-assisted scheme significantly outperformed conventional fixed-antenna NOMA systems.
  • vs. OMA: The proposed NOMA scheme achieved higher EE than Orthogonal Multiple Access (OMA) benchmarks.
  • vs. Discrete/Constrained Optimization: Continuous PA positioning yielded better EE than discrete or constrained positioning schemes.
• Scalability:
  • Antenna Count: EE increased with the number of PAs due to enhanced beamforming flexibility, though with diminishing returns due to circuit power consumption.
  • User Count: EE decreased as the number of users increased (due to interference and resource contention), but the proposed adaptive design maintained a performance gap over baseline schemes.

6. Significance

This paper represents a significant step toward 6G and B5G sustainable networks. By combining Pinching-Antenna technology with AI-driven resource allocation, the authors demonstrate a viable path to overcoming the limitations of fixed-antenna architectures in dynamic, energy-constrained IoT environments. The work proves that spatial adaptability, when coupled with intelligent learning algorithms, can drastically improve energy efficiency and spectral efficiency, addressing critical bottlenecks in future wireless networks.