🔢 mathematics

LLM-Enabled Automated Algorithm Design for Multiuser Fluid Antenna Communications

This paper proposes a novel LLM-enabled paradigm for the automated design of optimization algorithms, specifically enhancing genetic algorithms and creating a new heuristic called AutoPort, to efficiently solve the challenging port selection problem in multiuser fluid antenna communications and achieve near-optimal fairness performance.

Original authors: Gan Zheng, Fei Liu, Qingfu Zhang

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gan Zheng, Fei Liu, Qingfu Zhang

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

The Big Picture: A New Kind of Antenna

Imagine a traditional cell tower antenna as a rigid, fixed statue. It stands in one spot, facing one direction. To get a better signal, engineers usually have to build more statues (antennas), which is expensive and takes up a lot of space.

Fluid Antennas (FAS) are different. Think of them not as statues, but as a garden hose with hundreds of tiny nozzles along its length. You can't move the whole hose easily, but you can instantly switch which nozzle is spraying water. In a wireless system, this means the antenna can "jump" to different positions (ports) within a small space to find the clearest path for the signal.

The Problem: Too Many Choices

The challenge is that with hundreds of nozzles (ports) and many users (phones) trying to talk at once, figuring out which nozzle to use for which person is a massive puzzle.

The Goal: Make sure the user with the worst signal gets the best possible boost, so everyone gets a fair connection.
The Difficulty: There are so many combinations of nozzles to choose from that checking every single one (like trying every key on a giant keyring) would take longer than the age of the universe.
The Old Way: Engineers used to write "rules of thumb" (heuristics) by hand to guess the best nozzle. But these rules were often clunky, required a lot of human tweaking, and didn't work perfectly.

The Solution: Let an AI Write the Rules

Instead of an engineer writing the rules, the authors used a Large Language Model (LLM)—a super-smart AI that can read and write code—to automatically design the rules for solving this puzzle.

Think of it like this:

Old Way: A human coach writes a playbook for a football team, tries it, tweaks it, writes it again, and repeats for months.
New Way: You give the AI coach the goal ("Win the game") and the rules of football. The AI coach then invents, tests, and refines its own playbooks thousands of times in a few hours, discovering strategies a human might never think of.

How They Did It: Two Strategies

The paper tested two ways to let the AI design these rules:

1. The "Tweaker" Strategy (Optimizing Genetic Algorithms)
They started with a standard, well-known method called a "Genetic Algorithm" (which mimics evolution: you mix good solutions to make better ones).

What the AI did: Instead of changing the whole system, the AI acted like a mechanic. It looked at the two main "gears" of the machine—the Crossover (mixing solutions) and the Mutation (randomly changing solutions)—and rewrote the code for those gears to make them spin faster and smarter.
The Result: The AI found a way to mix and change the solutions that was much better than the standard human-designed way.

2. The "From Scratch" Strategy (AutoPort)
Here, they didn't give the AI any existing rules to tweak. They just said, "Here is the problem: Pick the best nozzles for these phones. Go make a program to solve it."

What the AI did: The AI invented a brand-new strategy from the ground up. It tried out ideas like "greedy approaches" (picking the best immediate option) and "simulated annealing" (cooling down a hot metal to find its strongest shape).
The Winner: The AI settled on a method called GRASP (Greedy Randomized Adaptive Search Procedure). It's like a treasure hunter who randomly picks a few promising spots to dig, but then carefully checks the surrounding area to make sure they didn't miss a bigger gold nugget nearby.

The Results: Beating the Competition

The authors ran simulations to see how well these AI-designed rules worked compared to old methods.

The "Exhaustive Search" (The Perfect Solution): This is the theoretical "gold standard" where you check every single possibility. It's too slow to actually use, but it sets the benchmark.
The "Deep Learning" (The Black Box): This is a popular AI method that usually needs huge amounts of data to learn. In this test, it performed poorly because the puzzle was too complex for it to learn quickly.
The "Human-Designed" (Basic GA): The standard method engineers usually use.
The "AI-Designed" (This Paper):
- The "Tweaker" version (GA-CM) got 99% of the way to the perfect solution.
- The "From Scratch" version (AutoPort) actually hit 100% of the perfect solution in many tests, finding the absolute best nozzle configuration.

Why This Matters

The paper claims this is a new way to solve hard engineering problems. Instead of humans spending years manually designing and tuning complex algorithms, an AI can do it automatically. The resulting "rules" are also written in clear code (Python), so humans can read them and understand why they work, unlike some other AI methods that are "black boxes."

In short: The paper shows that if you let a smart AI design the "rules of the game" for fluid antennas, it can find a winning strategy that is nearly perfect, beating both human experts and other types of AI.

Technical Summary: LLM-Enabled Automated Algorithm Design for Multiuser Fluid Antenna Communications

Problem Statement
Fluid Antenna Systems (FAS) offer a promising reconfigurable antenna architecture for 6G communications, capable of dynamically adjusting radiating element positions to provide spatial diversity within a constrained physical space. However, optimizing FAS involves a challenging large-scale combinatorial problem: selecting the optimal subset of antenna ports to serve multiple users while simultaneously designing beamforming vectors. The objective is to maximize the minimum Signal-to-Interference-plus-Noise Ratio (SINR) across all users under a total transmit power constraint.

Existing approaches face significant limitations. Manually designed heuristic algorithms are labor-intensive and often fail to achieve satisfactory performance. Deep learning methods, while promising, require extensive training data, high computational complexity for training, tedious parameter tuning, and suffer from poor generalization and lack of interpretability. Furthermore, traditional optimization techniques like alternating optimization incur high complexity and cannot guarantee global optimality.

Methodology
The paper proposes a novel paradigm leveraging Large Language Models (LLMs) for the automated design of heuristic algorithms (Automated Heuristic Design, or AHD). Instead of optimizing the variables (port selection) directly, the framework optimizes the algorithm itself. The authors utilize the LLM4AD platform and the Evolution of Heuristics (EoH) framework, which treats heuristics as code implementations that evolve through an iterative process of generation, evaluation, and selection.

The study investigates two distinct design strategies:

Optimization of Genetic Algorithm (GA) Operations:
- This strategy retains the basic structure of a standard Genetic Algorithm but uses the LLM to automatically design and refine specific operators: crossover and mutation.
- The LLM is prompted to generate new crossover and mutation functions based on a population of existing algorithms. These functions are evaluated on channel instances, and the best-performing variants are selected to evolve the next generation.
- The resulting optimized crossover function utilizes a frequency-based probabilistic inheritance strategy, while the optimized mutation function introduces diversity through element swapping and uniform random noise.
AutoPort: De Novo Heuristic Design:
- This strategy prompts the LLM to design a complete heuristic algorithm from scratch without being restricted to the GA family.
- The LLM is provided with a task description to select $n$ ports out of $N$ to maximize average SINR across multiple channel realizations.
- The evolutionary search discovers a Greedy Randomized Adaptive Search Procedure (GRASP) algorithm. This algorithm employs a constructive phase with probability-weighted random selection (based on channel gains) followed by a local search phase to refine the solution.

Key Contributions

First LLM-Enabled Framework for FAS: The paper presents the first application of LLM-driven automated algorithm design specifically for maximizing minimum SINR in multiuser FAS downlink systems. It automates the creation of customized, fully interpretable heuristics without requiring domain-specific hyper-heuristic tuning or pre-trained data.
Dual Design Strategies: The authors propose and validate two approaches: an incremental approach that enhances a baseline GA by optimizing its operators, and a discovery approach (AutoPort) that generates a new, standalone heuristic.
Interpretability and Generalization: Unlike "black-box" deep learning models, the resulting algorithms are human-readable Python code. The framework demonstrates strong generalization capabilities, performing well on unseen channel settings (e.g., Rician fading) and varying system parameters without retraining.

Simulation Results
Extensive numerical evaluations were conducted in a multiuser MISO-FAS downlink scenario with varying transmit powers, numbers of users ( $K=4, 8$ ), and channel models (Rayleigh and Rician).

Performance: The proposed methods significantly outperform conventional benchmarks.
- GA-CM (GA with LLM-optimized crossover and mutation) achieves over 99% of the optimal performance (obtained via exhaustive search for small-scale problems) and significantly outperforms the basic GA.
- AutoPort achieves near-optimal or optimal performance, often surpassing GA-CM.
- Both methods outperform traditional black-box optimization (Radial Basis Function), deep learning approaches (Transformer), and random selection. Notably, the Transformer-based method performed poorly, failing to extract features effectively for high-dimensional port selection.
Robustness: The LLM-designed algorithms maintained superior performance when tested on Rician channels, demonstrating robust generalization.
Scalability: As the number of ports increased, the performance of basic GA and random selection degraded due to increased correlation and complexity. In contrast, GA-CM and AutoPort improved or stabilized, verifying their effectiveness in large-scale scenarios.
Complexity: While AutoPort requires approximately twice the running time of GA-based methods (making it suitable for slow-varying channels), it remains computationally feasible compared to exhaustive search. The automated search process itself does not significantly increase the complexity of the final deployed algorithms.

Significance and Claims
The paper claims that this work represents a first step toward automated algorithm discovery in wireless communications. The significance lies in shifting the design paradigm from manual algorithm engineering to automated, LLM-driven search. The authors emphasize that their approach:

Reduces human effort and reliance on expert intuition.
Produces interpretable algorithms that offer insights into problem structures (e.g., the discovery of frequency-based inheritance or GRASP-like strategies).
Avoids the data hunger and generalization issues of deep learning.

The authors modestly note that this is an initial exploration. They identify future directions including jointly optimizing performance and complexity, addressing Channel State Information (CSI) estimation errors, and extending the framework to wideband channels (OFDM) and other reconfigurable antenna types (e.g., movable or rotatable antennas). They do not claim the method is a universal solution for all wireless problems but rather a promising new direction for solving difficult combinatorial optimization problems in the field.