This repository demonstrates how the Tsetlin Machine can be used for indoor localization using Bluetooth Low Energy (BLE) Received Signal Strength Indicator (RSSI) values. This example uses the BLE RSSI Dataset for Indoor Localization by Mehdi Mohammad on Kaggle.

This is a minimal example of indoor localization using the fingerprinting technique. In real-world applications, fingerprinting requires manually collecting RSSI measurements at various locations, which is a time-consuming process.

• The approach achieves an average localization error of ~2.1 meters.
• In practical scenarios, real-world accuracy may be significantly worse due to various environmental factors and dataset limitations.

This is an oversimplified example designed to demonstrate the use of the Tsetlin Machine for indoor localization. It does not represent a production-ready solution and has several limitations:

1. Dataset Limitations:

   • The exact methodology for collecting RSSI samples in the dataset is unclear.
   • It is assumed that the mobile device was accurately positioned within the predefined grid cells, but this is not explicitly verified.

2. Potentially Unrealistic Data Splitting:

   • The dataset is randomly split into 80% training and 20% testing, which is a common approach in machine learning.
   • However, this may lead to data leakage, where identical or near-identical samples appear in both the training and test sets, making the model's performance seem better than it would be in real-world deployment.
   • A more realistic approach might involve time-based splitting, but even then, there is a risk of overlap between training and test samples.

For a more in-depth discussion of challenges with indoor localization and potential improvements, see my master's thesis:
An Environment-Adaptive Approach for Indoor Localization Using the Tsetlin Machine.

In this project, we use multiple Regression Tsetlin Machines (RTMs)—one for each access point (AP). Each RTM predicts the distance between the mobile device and the AP. This method provides more accurate distance estimations compared to a simple path loss model, which often struggles with environmental variations.

After predicting distances, we apply a localization algorithm:

• Trilateration (supports exactly 3 access points)
• Least Squares Estimation (supports 3 or more access points, used in this project)

The Least Squares method refines the estimated position by minimizing the error between predicted and actual distances.

Before running the code, ensure you have the following:

1. Python (>=3.8) – Download from python.org
2. Tsetlin Machine – Install by following the instructions in the Tsetlin Machine Unified repository

This example also uses common third-party packages like matplotlib, pandas, numpy, and scikit-learn. These can be installed with pip by running:

pip install matplotlib pandas numpy scikit-learn

or by installing all dependencies via:

pip install -r requirements.txt

This example utilizes the BLE RSSI Dataset for Indoor Localization. The dataset is licensed under CC BY-NC-SA 4.0, meaning it cannot be used for commercial purposes.

⚠️ Dataset Not Included

Due to licensing restrictions, we do not distribute the dataset in this repository. You must download it manually from Kaggle and place it in the expected directory:

./data/iBeacon_RSSI_Labeled.csv

To run this example, run the following command:

python main.py

This script:

• Loads the dataset
• Trains a Regression Tsetlin Machine for each access point
• Uses the Least Squares algorithm for final position estimation
• Outputs localization accuracy statistics
• Plots two graphs

├── main.py              # Loads data, trains models, evaluates localization accuracy
├── least_squares.py     # Implements Least Squares method for position estimation
├── common.py            # Helper classes (Vector2D)
├── data/                # Directory where the dataset should be placed (not included)

The script outputs localization error statistics, including:

• Average Error (meters)
• Maximum/Minimum Error
• Standard Deviation of Error

Example output:

Average error: 2.10 m
Max error: 10.35 m
Min error: 0.15 m
Standard deviation: 1.22 m

Visualization:

1. RSSI vs. Actual Distance

Figure 1: This image illustrates how noisy the original RSSI data is. It is very difficult to determine the exact distance of a node based on RSSI alone, as the signal strength fluctuates significantly due to environmental factors such as interference, multipath effects, and obstacles.

2. Predicted vs. Actual Distance

Figure 2: After estimating distances using the Tsetlin Machine, we see that the predictions align somewhat with the actual distances. The points tend to follow the diagonal (y = x), which represents an ideal prediction. However, the predictions are not perfect, as there is still a noticeable spread around the optimal line, indicating estimation errors.

• This project is based on the Tsetlin Machine and its implementation in the Tsetlin Machine Unified (TMU) library.
• The dataset was created by Mehdi Mohammad and is available on Kaggle.
• The Least Squares localization implementation is based on the mathematical concepts described in:
  • Ye, Z., Xu, Y., Lin, J., Li, G., Geng, E., & Pang, Y. (2018).
    "An Improved Bluetooth Indoor Positioning Algorithm Based on RSSI and PSO-BPNN." Sensors, 18(9), 2820.
  • This is not a direct copy but an independent implementation of the mathematical method presented in the paper.

This project is for educational and research purposes only. The dataset is CC BY-NC-SA 4.0, which means:

• ✔ You can use and modify the dataset for non-commercial purposes.
• ✔ You must credit the original author when using or referencing the dataset.
• ❌ You cannot use the dataset or derivative works for commercial purposes.

For full details, see the CC BY-NC-SA 4.0 License.