⚡ electrical engineering

Genetic Algorithms and Reinforcement Learning Approaches for Joint Placement, Power Control, and Activation of Flying Base Stations

This paper evaluates Genetic Algorithms and Proximal Policy Optimization-based Reinforcement Learning for the joint 3D placement, power control, and activation of Flying Base Stations, revealing that while Genetic Algorithms achieve higher user connectivity in interference-limited scenarios, the RL approach offers superior power efficiency and interference reduction.

Original authors: Fady A. Abouelghit, Karim G. Seddik

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fady A. Abouelghit, Karim G. Seddik

Original paper licensed under CC BY 4.0 (https://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 city where the airwaves are crowded, and the ground-based cell towers that keep our phones connected are struggling to keep up. Sometimes, a sudden surge of people at a festival, a natural disaster that knocks out infrastructure, or simply a remote valley with poor coverage creates a gap in the network. To fill these gaps, engineers are looking to the sky. They propose using drones, or unmanned aerial vehicles, equipped with cell tower equipment to act as "flying base stations." These flying stations can hover over a crowd or a difficult terrain, beaming signals down to people on the ground. Because they are high up, they often have a clear, unobstructed line of sight to the users below, which is a distinct advantage over towers stuck on the ground. However, simply dropping a drone into the sky is not enough. The drone must be placed in the exact right spot, at the exact right height, and must broadcast its signal at the precise power level needed. If it is too high, the signal is weak; if it is too low, it gets blocked by buildings. If it broadcasts too loudly, it interferes with other signals; too quietly, and no one connects. Furthermore, the drone has a limited battery, so it cannot stay on forever. The challenge is to figure out the perfect combination of where to put the drone, how high to fly it, how loud to shout, and whether to turn it on at all, all while ensuring it does not disrupt the existing network.

In a recent study, researchers at the American University in Cairo tackled this complex puzzle by testing two different ways to solve it. They built a massive, realistic computer simulation of a wireless network that included both traditional ground towers and several flying drones. This simulation was sophisticated enough to mimic how radio waves actually travel through a city, bouncing off buildings and fading over distance, rather than using simple, idealized math. The goal was to maximize the number of people who could connect to the network while keeping the total energy used by the drones as low as possible. To find the best solution, the team pitted two distinct approaches against each other. The first approach was a "genetic algorithm," a method inspired by biological evolution. It works by creating a large group of possible solutions, testing them, and then mixing and mutating the best ones over many generations to see if they get better. The second approach was "reinforcement learning," a type of artificial intelligence where a computer agent learns by trial and error, receiving a reward for good decisions and a penalty for bad ones, gradually building a strategy for how to act in any given situation.

The researchers ran these two methods through various scenarios, changing the size of the city, the number of ground towers, and the number of drones available. They also tested two different goals: one that cared only about connecting as many people as possible, and another that tried to balance connectivity with saving energy. The results revealed a clear and surprising difference in how the two methods behaved. The genetic algorithm proved to be the more aggressive explorer. It consistently found ways to connect more people, especially in crowded or difficult environments where signals interfere with each other. It did this by actively moving the drones to new spots, adjusting their heights, and sometimes turning on multiple drones at once to cover more ground. In the most challenging test case, involving two ground towers and two drones, this method connected nearly twenty-four percent more users than the learning-based approach. It was willing to use more power to achieve this, treating the network as a dynamic system that could be reshaped to fit the needs of the users.

In contrast, the reinforcement learning agent took a much more cautious path. It learned to be conservative, often choosing to rely on the existing ground towers rather than risking the complexity of managing multiple drones. When the researchers added a penalty for using too much power, the learning agent responded by turning off one of the drones entirely and letting the ground towers do the heavy lifting. While this saved energy and reduced interference, it meant fewer people got connected. The learning agent seemed to prefer a stable, predictable solution over the high-risk, high-reward strategy of the genetic algorithm. It rarely turned on more than one drone, even when two were available, and it tended to keep the drones in positions that were safe but not necessarily optimal for maximum coverage.

The study did not claim that one method was universally superior, but rather that the choice depends entirely on what the network operator values most. If the priority is to ensure that every single person in a crowd gets a signal, the evolutionary, trial-and-error approach of the genetic algorithm is the better tool. It excels at finding complex, non-obvious arrangements of drones that maximize coverage. However, if the priority is to save battery life and keep the system simple and stable, the learning-based approach is more effective, even if it means some users might be left without a connection. The researchers noted that their findings were based on simulations, which are powerful tools for testing ideas before building real hardware, but they acknowledged that real-world conditions with moving people and unpredictable weather could change the outcome. Ultimately, the work provides a clear guide for engineers: you cannot have both maximum connectivity and minimum energy use with a single strategy. You must decide whether you want the network to be a bold, adaptive explorer or a careful, efficient guardian, and then choose the algorithm that matches that personality.

Technical Summary: Genetic Algorithms and Reinforcement Learning Approaches for Joint Placement, Power Control, and Activation of Flying Base Stations

1. Problem Statement

The paper addresses the optimization of Flying Base Stations (FBSs) in large-scale heterogeneous cellular networks. The core challenge is the joint optimization of three coupled decision variables for multiple FBSs:

Three-dimensional placement: Continuous horizontal coordinates ( $x, y$ ) and altitude ( $z$ ).
Transmit power: Continuous power levels ( $P$ ).
Activation status: Discrete on/off decisions ( $P_s \in \{0, 1\}$ ).

This problem is formulated as a Mixed-Integer Nonlinear Program (MINLP). The objective is to maximize a unified cost function that balances user connectivity (defined by Signal-to-Interference-plus-Noise Ratio, SINR, exceeding a threshold) and total FBS transmit power. The optimization must account for realistic propagation conditions, including air-to-ground and terrestrial links, LoS/NLoS transitions, and complex interference patterns in dense urban environments.

2. Methodology

The authors propose a unified evaluation framework to compare two distinct optimization paradigms under identical system models, constraints, and cost functions.

System Model and Channel Simulation

Network Topology: A heterogeneous network comprising fixed terrestrial Macro Base Stations (MBSs) and controllable FBSs serving a distributed set of users.
Channel Modeling: The study utilizes QuaDRiGa, a geometry-based stochastic channel simulator compliant with 3GPP standards. This replaces simplified analytical models to capture realistic path loss, shadowing, fading, and interference effects.
Connectivity Metric: A user is considered connected if their maximum achievable SINR (across all MBSs and active FBSs) exceeds a predefined threshold.

Optimization Framework

The objective function $J(d)$ combines normalized connectivity and normalized power consumption, shaped by hyperparameters ( $\beta, \gamma, \eta, w$ ) to allow tuning between connectivity maximization and power efficiency.

Two specific approaches are evaluated:

Genetic Algorithm (GA):
- Type: A real-coded, mixed-integer evolutionary algorithm.
- Mechanism: Chromosomes encode the full decision vector (position, power, and binary activation bit) for all FBSs.
- Operators: Uses BLX- $\alpha$ crossover for continuous variables and specific handling for binary activation bits (inheritance or random flipping). Mutation applies Gaussian noise to continuous genes and bit-flipping to activation bits.
- Strategy: Global search capability to escape local optima induced by the non-convex, mixed-integer nature of the problem.
Reinforcement Learning (RL):
- Type: A single-agent Proximal Policy Optimization (PPO) baseline.
- Mechanism: Treats the problem as a sequential decision-making process over a fixed horizon. The agent observes aggregate connectivity deltas and adjusts FBS parameters via continuous increments and activation signals.
- Strategy: Learns a policy to maximize cumulative discounted reward. The state space is limited to aggregate metrics (no per-user SINR maps) to ensure a fair comparison with the GA's feedback loop.

3. Key Contributions

The paper makes three primary contributions:

Large-Scale Heterogeneous Network Extension: The authors developed a simulation environment covering significantly larger geographical areas with multiple MBSs and FBSs, capturing denser interference interactions and diverse user distributions compared to typical small-scale studies.
Realistic Channel Modeling: Integration of the QuaDRiGa simulator ensures that propagation, fading, and interference are modeled with 3GPP-compliant accuracy, moving beyond simplified analytical approximations.
Unified Optimization Framework: The study establishes a common framework to directly compare evolutionary (GA) and policy-based (RL) methods for multi-FBS deployment. Crucially, it evaluates both methods under identical channel models, constraints, and cost function configurations, isolating the performance differences attributable to the optimization strategy itself.

4. Simulation Results

The study evaluates the approaches across two scenarios (varying area size and MBS count) and two cost configurations (connectivity-focused vs. power-aware).

Connectivity Performance: The Genetic Algorithm consistently outperforms the RL baseline in terms of total connected users.
- In interference-heavy, two-MBS scenarios, the GA achieved up to 23.9% higher connectivity than the RL agent.
- The GA effectively exploits spatial degrees of freedom, activating multiple FBSs and aggressively reshaping the network topology to serve more users.
Power Efficiency: The RL agent demonstrates superior power efficiency.
- The RL policy is more conservative, often deactivating redundant FBSs and relying more heavily on the always-on MBS tier.
- In several instances, the RL approach achieved lower power consumption per connected user (up to ~26% improvement in specific two-MBS settings).
Impact of Cost Shaping: Introducing a power penalty ( $\gamma > 0$ ) caused the RL agent to sharply reduce power and deactivate FBSs. The GA adapted by modifying spatial placement and service allocation rather than simply reducing power, maintaining higher connectivity at the cost of increased transmission power.
Activation Behavior: A key differentiator is the handling of binary activation. The GA frequently activates multiple FBSs to leverage combined coverage, whereas the RL agent tends to converge to single-FBS operation in power-aware settings, preferring simplified control strategies.

5. Significance and Claims

The paper claims that the choice of optimization method should be driven by specific operational priorities:

Genetic Algorithms are better suited for connectivity-maximizing deployments, particularly in interference-limited environments where spatial flexibility and the activation of multiple aerial nodes yield significant gains.
Reinforcement Learning is preferable when power efficiency and interference sensitivity are paramount, as the policy-based approach naturally converges to conservative, energy-saving configurations.

The authors emphasize that their primary contribution is not the superiority of one algorithm over the other in all contexts, but the unified framework that enables a fair, apples-to-apples comparison. They note that existing literature often decouples variables (e.g., fixing altitude) or uses simplified channel models, whereas this work addresses the fully coupled, continuous 3D placement–power problem with realistic channel modeling.

The study acknowledges limitations, including controlled assumptions regarding user distribution and the lack of user mobility in the simulation, suggesting that future work should explore more diverse operating conditions and user dynamics.