⚡ electrical engineering

A Hybrid Gauss Markov LSTM Mobility Model for Indoor OWC

This paper proposes a hybrid Gauss-Markov and LSTM mobility model that jointly predicts user position and device orientation to outperform conventional models in accuracy and communication stability for indoor optical wireless communication systems.

Original authors: Walter Zibusiso Ncube, Ahmad Adnan Qidan, Taisir El-Gorashi, Jaafar M. H. Elmirghani

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Walter Zibusiso Ncube, Ahmad Adnan Qidan, Taisir El-Gorashi, Jaafar M. H. Elmirghani

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 you are trying to have a high-speed conversation with a friend in a crowded room, but instead of using your voice, you are using a super-bright, invisible laser beam. This is Optical Wireless Communication (OWC). It's incredibly fast and can carry huge amounts of data, but it has one major weakness: it's like trying to talk through a straw. If you or your friend move even a little bit, or if you tilt your head, the "straw" gets misaligned, and the connection breaks or slows down.

To keep this connection alive, the system needs to know exactly where you are and which way your device is pointing before you actually get there. This is where the paper comes in.

The Problem: Predicting Human Chaos

The researchers wanted to build a system that could predict how people move indoors so it could keep the laser beam locked on target. They looked at two existing ways of guessing where people go:

The "Random Drunk" Model (Random Waypoint): Imagine a person walking who picks a random spot in the room, walks there, stops, picks another random spot, and repeats. This is easy to calculate, but it's terrible at predicting real humans. Real people don't walk in jerky, random jumps; they have momentum. They don't stop instantly, and they don't change direction without a reason.
The "Smooth Walker" Model (Gauss-Markov): This model is smarter. It knows that if you are walking north at 3 mph, you are likely to keep walking north at 3 mph for a little while. It accounts for inertia (the tendency to keep doing what you're doing). It's much better than the random model, but it still assumes humans move like robots on a track. It can't predict if you suddenly stop to tie your shoe, turn sharply to avoid a chair, or tilt your phone to take a selfie.

The Solution: The "Smart Co-Pilot" (Hybrid GM-LSTM)

The authors created a new, super-smart prediction system called Hybrid GM-LSTM. Think of it as a two-person team trying to guess your next move:

Person A (The Gauss-Markov Model): This is the experienced coach. They know the basics of physics. They say, "Based on your current speed and direction, you will likely be here in the next second." They provide a solid, smooth baseline prediction.
Person B (The LSTM/AI): This is the behavioral psychologist. They watch the "residual" (the difference) between what Person A predicted and what you actually did. They learn your specific habits. "Ah, every time Person A predicts you'll walk straight, you actually turn left because there's a coffee table there," or "You always tilt your phone down when you sit on the couch."

The Magic: The system combines them. It takes the smooth, physics-based guess from Person A and adds the "human behavior correction" from Person B.

Why Does This Matter?

The paper tested this new system against the old ones in a simulated room. Here is what they found, using simple analogies:

Accuracy: When the system tried to guess where you would be in the future, the new hybrid model was the most accurate. The old "Random" model was like guessing the weather by rolling dice. The "Smooth Walker" model was like a weatherman who only looks at the wind but ignores the clouds. The new model was like a meteorologist who uses both wind data and satellite images.
Speed: Even when people were walking fast (running around), the new model kept its cool. The old models got confused and made big mistakes, but the hybrid model adjusted quickly.
The Result (Data Rate): This is the most important part. Because the system knew exactly where your device was pointing, it could keep the laser beam perfectly aligned.
- With the old models, your internet speed would spike and crash like a rollercoaster (great for a second, then terrible).
- With the new model, your internet speed stayed smooth and steady, like a high-speed train on a straight track.

The Bottom Line

This paper proposes a way to make indoor laser internet much more reliable. By combining a simple math formula (which handles the physics of movement) with a smart AI (which learns human quirks and habits), they can predict where you are going and how you are holding your device. This ensures that your high-speed connection stays strong, even if you are walking, turning, or tilting your phone in a busy room.

In short: They taught the computer to stop guessing like a robot and start predicting like a human.

1. Problem Statement

Optical Wireless Communication (OWC) is a promising technology for high-capacity indoor networks due to its unregulated spectrum and high spatial reuse. However, OWC systems are fundamentally different from Radio Frequency (RF) systems because they rely on highly directional links (e.g., using VCSELs). Consequently, OWC performance is extremely sensitive to user mobility, specifically:

Spatial Position: Changes in user location alter the propagation distance.
Device Orientation: Changes in the angle of the receiver relative to the transmitter cause rapid fluctuations in channel gain.

Existing mobility models have significant limitations in this context:

Random Waypoint (RWP): Assumes memoryless, independent movement steps. It fails to capture the inertia and continuity of human motion, leading to poor channel prediction.
Gauss–Markov (GM): Introduces temporal correlation for smoother trajectories but remains a linear stochastic model. It cannot accurately capture complex, non-linear human behaviors (e.g., abrupt turns) or dynamic device orientation changes.
Pure Machine Learning Models: While capable of learning non-linear patterns, they often lack robustness in unseen scenarios and require massive real-world datasets for training.

The Core Challenge: There is a need for a mobility model that accurately predicts both user position and device orientation to enable reliable channel estimation and stable data rates in dynamic indoor OWC environments.

2. Methodology

The authors propose a Hybrid Gauss–Markov–Long Short-Term Memory (GM–LSTM) model. This approach combines a physics-based stochastic model with a data-driven deep learning component to predict the user state vector $s_u(t) = [x, y, v, \psi, \theta, \phi]^T$ (position, speed, direction, and orientation angles).

The prediction for the next time step $t+1$ is formulated as:
$\hat{s}_u(t+1) = f_{GM}(s_u(t)) + f_{LSTM}(s_u(t-k:t))$

A. The Gauss–Markov (GM) Component

Role: Acts as a structured baseline.
Function: It captures the temporal correlation and inertia of human motion. It ensures smooth transitions in position, velocity, and direction based on previous states.
Limitation: It assumes linear dynamics and cannot model complex behavioral deviations.

B. The LSTM Component

Role: Learns the residual behavior.
Function: Instead of predicting the full state, the LSTM is trained to predict the difference (residual) between the ground-truth mobility state and the GM prediction.
Input: A sequence of past mobility states (including position and orientation) over a window of length $k$ .
Output: A correction term that accounts for non-linear patterns, abrupt direction changes, and device orientation variations that the GM model misses.

C. Training Strategy

Data Generation: Synthetic ground-truth trajectories are generated using a GM-based simulator with bounded behavioral variations (simulating realistic indoor movement).
Supervised Learning: The LSTM is trained offline to minimize the error between the GM prediction + LSTM residual and the actual ground-truth state.
Inference: During operation, the GM model provides the initial estimate, and the trained LSTM refines it using recent history.

3. Key Contributions

Novel Hybrid Architecture: The first integration of a Gauss–Markov stochastic model with an LSTM to jointly predict user position and device orientation for indoor OWC.
Residual Learning Approach: By having the LSTM learn only the residual (deviation) from the GM model, the system leverages the physical constraints of motion while capturing complex non-linear behaviors, reducing the data requirements compared to pure deep learning models.
Joint State Prediction: Unlike previous models that focus solely on location, this model explicitly handles orientation dynamics, which is critical for directional OWC links.
Comprehensive Evaluation: The model is evaluated not just on trajectory accuracy, but on its direct impact on channel estimation and per-user data rate stability.

4. Results and Performance Evaluation

The proposed model was tested in a $5\times5\times3$ m $^3$ indoor environment with VCSEL-based APs. It was compared against Random Waypoint (RWP) and standard Gauss–Markov (GM) models across three metrics:

Channel Prediction RMSE vs. Prediction Horizon:
- As the prediction horizon increased (up to 1000 ms), all models showed increased error.
- RWP had the highest error due to its memoryless nature.
- GM performed better than RWP due to temporal correlation.
- GM–LSTM achieved the lowest RMSE, demonstrating superior accuracy in predicting future channel states over time.
Channel Prediction RMSE vs. User Speed:
- Error increased for all models as user speed rose from 0.2 m/s to 1.6 m/s.
- GM–LSTM maintained the most robust performance across all speeds, effectively compensating for the rapid channel variations caused by faster movement.
Data Rate Evolution:
- RWP caused significant fluctuations, spikes, and deep drops in data rate due to inaccurate link adaptation.
- GM provided smoother variations but still suffered from degradation during complex maneuvers.
- GM–LSTM delivered the most stable data rate, staying consistently close to the maximum achievable throughput. It significantly reduced performance variability, which is crucial for Quality of Service (QoS).

5. Significance and Conclusion

The paper demonstrates that integrating model-based physics (GM) with data-driven learning (LSTM) creates a superior framework for OWC mobility management.

Reliability: The hybrid model significantly reduces channel prediction errors, leading to more reliable link adaptation and handover decisions.
Stability: By accurately predicting orientation and position, the system minimizes data rate fluctuations, ensuring a stable user experience in dynamic indoor environments.
Future Impact: This approach paves the way for adaptive resource allocation and mobility-aware handover strategies in next-generation LiFi and high-frequency wireless networks. Future work will extend this to complex scenarios involving user blockage and multipath propagation.