NAVTEK-AUTO, MACHINE LEARNING

Petri Välisuo
University of Vaasa

2025-03-12
NAVIGATION TECHNOLOGIES FOR
AUTONOMOUS SYSTEMS
.
Machine Learning (Petri Välisuo)
.

.
CONTENTS
• AI and Machine Learning
• Classi�cation
• Regression
• Reinforcement learning
• Applications in navigation
AI AND MACHINE LEARNING
Arti�cial intelligence (AI), the ability
of an articat (like robot) to perform
tasks commonly associated with
intelligent beings, such as the ability
to reason, discover meaning,
RL generalize, or learn from past
experience.

• AI is often related to human like or

rational thinking or acting like human.
• Rational reaction is somethint which is
Figure 1: AI, machine learning and deep learning
needed to achieve the the best expected
outcome in case of uncertainty.
• RL = Reinforcement Learning
• LLM = Large Language Model
MACHINE LEARNING
• The research on machine learning methods for positioning is becoming increasingly
popular due to the increased calculation capacity of positioning hardware and
advances in ML methods.
• Machine learning methods are used in positioning when the assumptions of
standard multilateration techniques are not satis�ed; for example, when the
measurement noise is long-tailed or skewed or when some measurements are
outliers.
• ML methods may allow increasing robustness by excluding outliers, integrating
heterogeneous information sources, and utilizing ad hoc observations not initially
intended for positioning.
METHODS
• ML methods are combination of mathematical methods, statistics and computer
science.
• Mathematical methods include, Least Mean Squares, Gradient Descent, Gauss-
Newton, Levenberg-Marquardt optimisation methods, discriminant analysis, etc.
• Statistical methods include hypothesis testing, Bayesian �ltering, etc.
• Some of the methods include heuristic and metaheuristic methods, which uses
knowledge of the problem to �nd suf�ciently good solutions. These are for example
Particle Swarm Optimisation (PSO), Genetic Algorithms (GA), etc.
WHAT IS MACHINE LEARNING?
• If y is continuous, the ML is regression
(not necessarily linear)
• if y is discrete, the ML is classi�cation
• The problem is to �nd optimal function
f() and model parametes θ to minimize
the prediction loss L.
• The problem is easily solved if the
dimensionality, p, of x̄ is small.

⎛⎡
Figure 2: Training of Machine Learning model

⎡ ⎤
y1
x11 x12 x13 … x1p
y^ = f(x̄, θ)
⎢ y2 ⎥ ⎜⎢ x21
⎢ ⎥=f⎜ ⎢
x22 x23 … x2p
⎢ ⎥ ⎜⎢
⎢ ⋮ ⎥ ⎜⎢
⎜⎢ ⋮
⎣y ⎦
⋮ ⋮ ⋱ ⋮
n
⎝⎣ x xn2 xn3 … xnp
n1
NN TRAINING

X yh L

Error

dError
Update
dE
network

Figure 3: Training of Deep Neural Network

CLASSIFICATION
SUPPORT VECTOR MACHINE

Figure 4: Classi�cation with Support Vector Machine

OUTLIER DETECTION
Isolation forest

Figure 5: Outlier detection with Isolation Forest

REGRESSION
ROBUST REGRESSION
Estimated coefficients (true, linear regression, RANSAC):
82.1903908407869 [54.17236387] [82.08533159]

Figure 6: Robust RANSAC regressor example

NON-LINEAR REGRESSION

Figure 7: Non-linear regression with SVR: Fitting curve to sinusoidal data

REINFORCEMENT LEARNING
• Learning process: Agent learns by interacting with the environment by taking
actions and receiving feedback. It tries to learn a policy that maximizes the
cumulative reward.
• Feedback: Instead of receiving feedback for every sample, it receives feedback
based on actions.
• Objecive: Learn an optimal strategy (policy)
• Does not need labeled data
• For example it can learn to play a game by trial and error
Figure 8: Navigation using Deep Reinforcement Learning (Zhu and Zhang 2021)
Figure 9: Application scenarios for DRL based navigation (Zhu and Zhang 2021)
USE CASES
USE CASES OF ML IN GNSS NAVIGATION
• Veri�cation of signal acquisition. The raw signal from the satellite is examined and
missin or interfered signal can be detected.
• Non-Line-of-Sight (NLOS) situation detection
• Multipath propagation situation detection
• Signal blockage detection
• Indoors-outdoors detection, and positioning system switching
• GNSS anomaly detection
• Atmosphere modelling
• GNSS security: Spoo�ng, jamming and meaconing attack detection
• Satellite selection to optimize for Geometric Dilution of Precision (GDOP) and
minimize errors by excluding problematic satellites
• Satellite orbit determination
• Position �ngerprinting
NON-LINE-OF-SIGHT AND MULTIPATH PROPAGATION

Legend
LOSand NLOSin positioning
Symbol Count Description
1 User (tag) Non Line Of Sight

4 anchors Line Of Sight

1 Smart phone

Figure 10: Non-Line-of-Sight and multipath propagation (Elsanhoury et al. 2022)

SOME EXAMPLES OF ML IN POSITIONING
• Robust positioning ML-based variable selection algorithms to remove outlying
observations (Liu et al. 2019).
• Adaptive positioning Change the model parameters according to the conditions. For
example, the positioning method or observables can be switched when an agent
moves outdoors to indoors or when an NLOS situation is detected [Xia and
Weitnauer (2019)](Carrera Villacrés et al. 2019).
• Collaborative positioning by sharing the information between agents. The
conditions of standard positioning methods are not satis�ed due to skewed noise
distribution and non-convexity. ML-based methods, such as cognitive PF, may help in
solution �nding (Minetto, Gurrieri, and Dovis 2020).

continue…
Examples of ML in positioning continued

• Model free positioning ML learns the model from the problem domain, thus
allowing model-free positioning. Usefull particularly when utilizing ad hoc
information. Examples are Wi-Fi �ngerprinting-based positioning (Li et al. 2019)
utilizing, for example, nonlinear ensemble regression methods (Marthala 2020).
• Signals of Opportunity Intelligent variable selection from heterogeneous
information sources with ML can be helpful when utilizing ad hoc information
sources, but also within a standard positioning infrastructure with NLOS problems
(Vandermeeren et al. 2018). For example, nearest-neighbors, k-means, and support
vector machines and outlier detection methods are often-used algorithms for
feature selection.
MACHINE LEARNING METHODS USED IN POSITIONING
• Decision trees (DTs)
• Random Forests (RF)
• Regression analysis (Linear/Logistic)
• K-Nearest Neighbour (KNN)
• K-means clustering
• Naive Bayes (NB)
• Extreme Learning Machine (ELM)
• Long-short Term Memory (LSTM)
• Gaussian Process Regression (GPR)
• Support Vector Machine (SVM)
• Neural Networks (NN)

(Siemuri et al. 2022)

THE POPULARITY OF ML METHODS

Figure 11: Popularity of ML models (Siemuri et al. 2022)

THE POPULARITY OF DEEP LEARNING MODELS

Figure 12: Popularity of DL models (Siemuri et al. 2022)

REFERENCES
Carrera Villacrés, José Luis, Zhongliang Zhao, Torsten Braun, and Zan Li. 2019. “A
Particle Filter-Based Reinforcement Learning Approach for Reliable Wireless
Indoor Positioning.” IEEE Journal on Selected Areas in Communications 37 (11):
2457–73. https://doi.org/10.1109/JSAC.2019.2933886.
Elsanhoury, Mahmoud, Petteri Makela, Janne Koljonen, Petri Valisuo, Ahm
Shamsuzzoha, Timo Mantere, Mohammed Elmusrati, and Heidi Kuusniemi.
2022. “Precision Positioning for Smart Logistics Using Ultra-Wideband
Technology-Based Indoor Navigation: A Review.” IEEE Access, 1–1. https://
doi.org/10.1109/access.2022.3169267.
Li, Lin, Xiansheng Guo, Nirwan Ansari, and Huiyong Li. 2019. “A Hybrid Fingerprint
Quality Evaluation Model for WiFi Localization.” IEEE Internet of Things Journal 6
(6): 9829–40. https://doi.org/10.1109/JIOT.2019.2932464.
Liu, Wanke, Jianlong Li, Qi Zeng, Fei Guo, Renpan Wu, and Xiaohong Zhang. 2019.
“An Improved Robust Kalman Filtering Strategy for GNSS Kinematic
Positioning Considering Small Cycle Slips.” Advances in Space Research, Multi-
GNSS: Methods, Bene�ts, Challenges, and Geosciences Applications, 63 (9):
2724–34. https://doi.org/10.1016/j.asr.2017.11.041.
Marthala, Raja Vardhan Reddy. 2020. “Wi� Based Indoor Positioning - a Machine
 Learning Approach.” Master’s thesis, University of Vaasa; University of Vaasa.
https://osuva.uwasa.�/handle/10024/11126.
Minetto, Alex, Alessandro Gurrieri, and Fabio Dovis. 2020. “A Cognitive Particle
Filter for Collaborative DGNSS Positioning.” IEEE Access 8: 194765–79. https://
doi.org/10.1109/ACCESS.2020.3033626.
Siemuri, Akpojoto, Kannan Selvan, Heidi Kuusniemi, Petri Valisuo, and Mohammed
E. Elmusrati. 2022. “A Systematic Review of Machine Learning Techniques for
GNSS Use Cases.” IEEE Transactions on Aerospace and Electronic Systems, 1–42.
https://doi.org/10.1109/TAES.2022.3219366.
Vandermeeren, Stef, Samuel Van de Velde, Herwig Bruneel, and Heidi Steendam.
2018. “A Feature Ranking and Selection Algorithm for Machine Learning-Based
Step Counters.” IEEE Sensors Journal 18 (8): 3255–65. https://doi.org/10.1109/
JSEN.2018.2807246.
Xia, Nan, and Mary Ann Weitnauer. 2019. “TDOA-Based Mobile Localization Using
Particle Filter With Multiple Motion and Channel Models.” IEEE Access 7:
21057–66. https://doi.org/10.1109/ACCESS.2019.2897936.
Zhu, Kai, and Tao Zhang. 2021. “Deep Reinforcement Learning Based Mobile Robot
Navigation: A Review.” Tsinghua Science and Technology 26 (5): 674–91. https://
doi.org/10.26599/TST.2021.9010012.